The argument in one line.
Agentic engineering replaces vibe coding by keeping humans in control as architects while AI generates 95% of code through harnesses, context tools, service layers, and autonomous review loops that enforce clean structure and measurable outcomes.
Read if. Skip if.
READ IF YOU ARE…
- A developer with 3+ years of experience shipping production code who wants to understand how to architect AI agents into their existing workflow without losing control.
- An engineer currently using Cursor or similar AI coding tools who feels like they're 'vibe coding' and wants a systematic framework to scale from 30% to 90% AI-generated output.
- A technical leader or senior dev evaluating whether agentic engineering is worth the infrastructure investment for their team in 2025-2026.
SKIP IF…
- You're new to professional development or still learning fundamentals — this assumes you ship production code and understand architecture patterns.
- You're building in a domain with hard regulatory or compliance constraints that require human code review at every step — the autonomous PR loop won't work for you.
- You're philosophically opposed to AI code generation or looking for arguments against it — this is a how-to for believers, not a debate.
The full version, fast.
Vibe coding offloads thinking to the model and produces brittle, bloated code; agentic engineering keeps you as architect while the model does the typing. The method is a stack working in concert: a strong harness like Cursor paired with a top-tier model, the opensrc tool to pull real source code from packages into the repo so the agent reads canonical truth instead of hallucinated docs, a code-structure skill that consolidates duplicated logic into a service layer after each feature, and a greploop that iterates pull requests against an automated reviewer until they earn a five-out-of-five score. Keep features small, contexts lean, and PRs minimal. Ship early, stay delusional, and let tight feedback loops handle correctness.
Members feature
Chat with this breakdown.
Modern Creator members can chat with any breakdown — ask for the hook, quote a framework, find the exact transcript moment. Unlocks at T2: refer 3 friends + add your own API key.
Create a free account →Who's talking.
00:00hostDavid Ondrej
00:21guestMicky (Michael Shimeles)
Where the time goes.
00:00 – 03:05
01 · What is a harness?
Harness vs model: model predicts next token; harness wraps it with tools, system prompts, markdown files. Cursor above Claude Code above Codex. GPT-5.5 xHigh Fast for logic, Opus 4.7 Max for UI.
03:05 – 10:20
02 · opensrc: replace docs with the actual repo
Vercel opensrc tool dumps any package source into repos/ folder. Reference codebase in prompts. Agent finds exact functions. Context stays lean via search tools, no RAG needed.
10:20 – 16:55
03 · Code structure skill: service layer
After every feature run a service-layer refactor skill. Reusable functions, no duplicated logic. Clean codebase means tight context for next agent session.
16:55 – 22:00
04 · greploop: autonomous PR review until 5/5
Push PR, run /greploop, agent reads PR, gets Greptile confidence score 1-5, fixes feedback, loops until 5/5. Ships 9/10 PRs autonomously.
22:00 – 27:10
05 · Stack choices optimized for agents
Svelte over React (HTML+TS, no footguns). Convex over Supabase (backend is pure TS code, no dashboard). Open-source libraries: give agent the repo.
27:10 – 34:05
06 · Mindset and model investment
Best-in-class models matter. 200 per month subsidized. Autonomous greploop runs prove the ROI. Claude used to negotiate a contract 3x higher.
34:05 – 39:30
07 · Launch early, be delusional
SF founders launch half-broken MVPs, raise millions, ship v2. You are still fixing one more feature. Delusion plus belief equals launch equals real feedback.
39:30 – 45:05
08 · Cybersecurity in the agentic era
Passphrase for voice clones, 2FA via Authenticator not SMS, password manager, 14-day package rule, paste breach tweet into Claude to check exposure.
45:05 – 53:52
09 · Future: knowledge work boom
More bullish on knowledge work than agentic engineering. Tooling lags models. Non-technical is no longer an excuse. Field resets every six months.
Lines worth screenshotting.
- Agentic engineering keeps the human as the architect while the AI writes 95% of the code — vibe coding hands that thinking to the agent entirely.
- The model is just a next-token predictor; the harness wrapping it — APIs, tools, system prompts, agent.md files — is what actually makes it useful.
- GPT-5.5 Extra High outperforms Claude on large, complex codebases while Opus 4.7 wins on UI and frontend changes — use both, swapping by task type.
- Cursor benchmarks higher than Claude Code and Codex for their own models, which surprises most developers who assume the native harness is always best.
- The opensrc context tool fetches package source code directly into the codebase, giving the agent the single best source of truth rather than handwritten docs.
- Harnesses are getting thinner as models improve — modern agents can infer the tech stack from the code itself, making many CLAUDE.md entries redundant.
- The agent.md should only document what the model wouldn't know on its own: vision, specific conventions, non-obvious project structure — not React or TypeScript.
- Greptile's greploop runs autonomous 5/5 PR reviews and is the most underused tool in the senior agentic engineering stack.
- A post-feature code-structure skill that rewrites the architecture after each build keeps the codebase clean as AI-generated code compounds.
- Running multiple agent harnesses in parallel on different parts of a codebase is the real productivity unlock — not prompting a single agent faster.
- The distinction between vibe coding and agentic engineering is not the tools used but who holds the architectural decision-making authority.
- Investing in a great harness returns more value than chasing the latest model, because the harness determines how much of the model's capability actually reaches your output.
Steal the loop, not just the stack.
Agentic engineering playbook
The greploop is the unlock: an autonomous PR review cycle that runs until the agent earns a 5/5 review score before merging, and it only works because of small PRs and clean service layers.
- Add opensrc to your harness: fetch any thin-docs library repo into your codebase folder and reference it in every prompt instead of guessing from human-written docs.
- Run a code-structure pass after every feature, not for the human, for the next agent session picking it up cold.
- Keep PRs minimal so the review loop can close cleanly; a 100-line PR with a 5/5 score ships faster than a 5000-line PR you are hand-reviewing.
- Pick frameworks where the syntax is familiar model territory (Svelte, not React hooks) and backends where everything is code (Convex, not a GUI dashboard).
- Add the 14-day package rule to your CLAUDE.md so the agent never installs a fresh supply-chain attack vector.
- The field resets every six months; catching up is cheap if you treat non-technical as temporary, not a fixed identity.
Terms worth knowing.
- Agentic engineering
- A disciplined approach to building software where the human directs AI agents as architect and reviewer, providing tight context, plans, and feedback loops rather than handing the entire thinking process over to the model.
- Vibe coding
- A loose, intuition-driven style of building with AI where the developer offloads most thinking to the agent and accepts whatever it produces, with little structured planning or review.
- Harness
- The software wrapper around an AI model — APIs, tools, system prompts, and instruction files — that lets the model take actions like reading files, running commands, or editing code.
- Cursor
- An AI-first code editor that wraps multiple frontier models in a single harness, letting developers switch between providers like Anthropic and OpenAI inside one workflow.
- Claude Code
- Anthropic's command-line coding agent that runs in a terminal, reads and edits files, and executes shell commands as part of its tool set.
- Codex
- OpenAI's coding agent product that runs GPT models against a developer's codebase, available through a subscription that heavily subsidizes usage.
- Tool call
- A structured request an AI agent makes through its harness — like reading a file, running a search, or executing a command — to take action outside of plain text generation.
- Token
- The small unit of text — a word, fragment, or character — that language models read and produce. Models predict the next token in a sequence rather than thinking in whole sentences.
- Context window
- The maximum amount of text an AI model can hold in working memory during a single session. Once it fills up, the model loses earlier information and quality degrades.
- Context engineering
- The practice of carefully shaping what information an AI agent sees — keeping prompts focused, supplying the right source material, and avoiding bloat — to maximize output quality.
- opensrc
- An open-source CLI tool from Vercel that downloads the full source code of a package or repository into your project so an AI agent can reference real code instead of human-written documentation.
- agents.md
- A markdown file placed in a project root to give AI coding agents standing instructions — vision, conventions, or non-obvious rules — that they read before working on the codebase.
- RAG (retrieval-augmented generation)
- A technique where relevant snippets are pulled from a vector database and injected into an AI prompt so the model can answer with knowledge it wasn't trained on.
- Vector indexing
- Converting code or text into numerical embeddings stored in a database so an AI can later search by meaning. Modern agents increasingly skip this in favor of direct file search.
- Slash compact
- A Claude Code command that summarizes a long conversation to free up context window space, letting the session continue without starting fresh.
- Service layer
- A code architecture pattern that consolidates shared functionality into reusable modules so the same logic isn't rewritten across many files — keeping a codebase easier for both humans and agents to navigate.
- Code smell
- Surface-level signs that a codebase is decaying — duplicated logic, sprawling files, tangled dependencies — that make future changes harder even when the code still runs.
- Skill
- A reusable instruction package — typically a markdown file plus optional scripts — that an AI coding agent can invoke on demand to perform a specific structured task, like reviewing or refactoring code.
- Greptile
- An AI code review service that inspects pull requests on GitHub and returns scored feedback, flagging issues with a confidence rating from one to five.
- Pull request (PR)
- A proposed set of code changes submitted for review before being merged into the main project, typically through GitHub or a similar platform.
Things they pointed at.
16:30conceptKarpathy auto-research loop
16:30toolMatt Pocock code structure skill
Lines you could clip.
09:00
“Vibe coding offshores the thinking to the agent. Agentic engineering: you do the thinking, then let your minions do the work.”
13:05
“The model is a really dumb person with photographic memory.”
25:20
“Code is the single best source of truth. Human-written docs are the worst.”
35:10
“That belief is why they raised ten million and we are still fixing one more feature.”
42:00
“Never install a package younger than fourteen days.”
52:00
“Even our thinking changes every six months. Last year we were stuffing the agent with context. Now we do the complete opposite.”
Where the conversation goes.
00:00 – 03:05denseHarness and model selection
03:05 – 10:20denseContext engineering with opensrc
10:20 – 16:55denseCode structure and service layers
16:55 – 22:00densegreploop autonomous PR review
22:00 – 27:10steadyTech stack Svelte and Convex
27:10 – 34:05steadyMindset and model investment
34:05 – 39:30steadyLaunch early and delusion
39:30 – 45:05denseCybersecurity in agentic era
45:05 – 53:52steadyFuture of knowledge work
Word for word.
metaphoranalogystory
00:00Agentic engineering is the future. In 2026, the people who ship 100 x faster aren't typing prompts into a chatbot.
00:06They're running multiple agent harnesses in parallel. So I had Mickey on the podcast, a senior developer who has AI write 95% of his code, to walk through his exact AI stack.
00:17The tools, the models, the loops. This is the David Andre podcast. Enjoy.
00:22RadMikki, how do you think about building with AI in 2026?
00:25Yeah. It's quite changed since the last time we talked. It's no longer vibes.
00:29Like, we gotta be serious with this stuff. But at the same time, I'll be honest, like, I think I would say in the last three months, all the code that I've created, I would say 95% was generated by AI. And you're an actual developer.
00:41Right? So, like, let's Yeah. Like, I'm an actual engineer developer.
00:43Like and, like, there's a part of me that misses writing code for fun. I'll do it every now and then in the weekends. But, like, I think it'd be foolish to not see where we're going.
00:52Yeah. The models are not perfect. They're at a point where, like, there are productivity gains, if you understand the vertical that you're in.
00:5990% of the applications people are building. A little bit of brains and AI will take you a long way. So I'm gonna share, like, the tools I use in my workflow, and I think anybody else can do this.
01:08It's super replicatable. I'm actually gonna give the exact tools, and it's also gonna feel like Carpathi's, like, auto research loop. So I think it'll be fun.
01:15Awesome. Let's jump into it. First is harness.
01:18I chase the model. Right? I think the model is the more important thing, but I use cursor.
01:23And And right now, I'm using GBT 5.5.
01:25Interesting. Okay. So that's funny.
01:27I'm using Opus 4.7 max fast inside of cursor, but also inside of cursor agent. Here's actually a big difference a lot of people don't know. There have been benchmarks.
01:36Right? And, like, these are, like, benchmarks done by, like, actual developers where cursor performs both Cloud Code and Codex for their models.
01:45Right? So That's crazy. A lot of people have been dunking on Opus 4.5 saying Opus 4.5 is terrible.
01:50Some people on Twitter call it Slopus. It's still a good model, especially when it comes to UI changes. So I will preface there whenever I do anything UI or front end, uh, wise.
01:59I will use Opus 4.7, but I'm using a max. I have no time for any other variant. But the harness matters.
02:05Now a lot of people are priced out of Cursor because Cursor doesn't subsidize the way Codex and Claude does. Cursor, in my opinion, is the best harness. I can switch between models.
02:14Their new agentic view is pretty, pretty nice. So I guess let's touch on the harness. Right?
02:19Because a lot of people, they say, word that's becoming more popular now in in the last two, three months. Obviously, it's been around for longer, but a lot of people it's their first time hearing this really, and they don't really understand it fully. So you have the model.
02:29What is the harness? Right? It's everything around it.
02:30So believe it or not, the model itself can't do anything. The model is just a predictor of next text. Technically, the model doesn't even think.
02:39Whatever English you give it, it converts it to tokens. Those tokens are mapped on a graph. And then what it does is it looks at that graph and it predicts the next token.
02:48Some cool mathematics happens there, and then it returns a token to you, which is converted to English. So the model doesn't think. The model doesn't do.
02:56The model just predicts the next text. What a harness does, you can think of a harness as a wrapper of a model. And this wrapper is APIs, tools, specific system prompt, markdown files like agent MD files.
03:10These are all things that wrap the model that guide it to perform or specialize in a specific action. Real quick, if you want everything mentioned in this video completely for free, click the first link below the video, and I will email you all the skills, all the repos, all the tools mentioned in this podcast. Again, it's completely for free.
03:27So go blow the video and grab it now. Now the main important thing with the cursors and the clon code and the codex is the tools it gives it. Whenever you use cursor
03:35or any sort of agent, you you notice in the trace and the response it gives you, it says, oh, it read this file. It searched this thing. Right?
03:42Those are tool calls that it's making. Again, the agent doesn't have the ability to do so, but the harness gives it the tool so that the agent can do this. The models have got to a point where an investment in a really good harness will maximize the output you get from the model.
03:56And we know this because people's experience with Cloud Code and Cursor is not the same. Yeah. As long as people use the best model available.
04:03And that's the thing. Right? The model does matter.
04:06Um, I will GBT 5.5 extra high, especially I work mainly on large code bases, so I may be in that bias state. It's just really good at understanding large code bases, complex code bases.
04:18And it's been two weeks, and other than UI changes, this is the only model I've been using. There are some tools that I highly recommend. The first one is open source.
04:27Now it it sounds like the term open source, but it's actually a repo. If you go on Google, it's actually by Vercel.
04:34Shout out to Vercel for open sourcing this. Basically, what open source does is it fetches the source code of whatever package you're working on and dumps it into the code base. And I'll give a specific example.
04:45Like, this is a pretty large code base of an app I'm working on. The structure doesn't really matter. What matters is I have this grayed out folder here that says open source, a folder called repos.
04:55If I click on that, there's a folder called github.com. Where am I? See a bunch of popular projects or companies you may have heard heard of.
05:02Browser use, Composeo, Daytona, OpenClaw.
05:07A lot of the technology is actually open source, which is fantastic for agents. Right? Instead of dumping in your docs or what the docs or whatever that are man made, you can give the agents the best source of truth, which is the code.
05:19Because the code is the single best source of truth. What I've done is in an agents dot m d now I'm not a big fan of agents dot m d files Interesting. Unless it's something that the agent wouldn't know by hand.
05:32For example, like, people put, oh, this is a React code base on the agent dot m d file. And I believe this would have been a good idea if we were using Sonnet fourth, you know, four four four five or, like, Opus four o. But, again, the models have gone so good that it will just read your code base and know exactly what the tech stack is and stuff like.
05:52So if you notice, uh, the the harnesses are sort of getting lighter and lighter and thinner and thinner. This is why pie has been winning so much. You actually don't need much.
05:59Like, the models are really, really good. So you just wanna tell it the stuff that's not obvious. Right?
06:03Like, how this project will be used? What's the vision behind it? Exactly.
06:07Like, if there's a certain structure that I have and stuff like that. Now in this case, I tell it that it can fetch for additional source code. And by the way, all this is AI generated.
06:16I I don't handwrite this. Right? So if anyone reads this, it's like, oh, so smart.
06:20I wish. AI generated this. It's like, I think, and I tell the AI what to do.
06:26And, literally, there's a small block block that says to fetch source code for a package or repo, you need to understand and run. So I tell it basically how to add packages. And the best way to add package, can, like, maybe show an example.
06:38Actually, we'll just do open source itself. So let's say I'm building a product that uses this package or this open source repo or you're using something like browser use. You find the repo, and then you can literally go to terminal, and then I can do MPX open source, and then just paste the repo.
06:58And then watch this. Okay. This failed.
07:01Why did this fail? Usage. Oh, because I'm on a Linux machine.
07:05Okay. Um, don't mind that. But, literally, all I have to do is just run this command and what's going to happen, and then it will pop up here.
07:13Now here's the cool part, uh, David. Whenever I want to build a feature that uses set technology, I will tell it in the prompt, reference the code base.
07:23Right? So I'll go back to my drawing board here. Every time I prompt, let's like, for this project, I'm using Daytona a lot, is a really powerful sandboxing platform.
07:33There's a lot of cool stuff you can do. It's very, very technical. It's very, very deep.
07:38I can go in and spend some time, but I wanna build fast. I wanna ship fast. Right?
07:42My competitors are shipping at light speed. So in my prompt, I will tag the folder and I will say reference the code base. Is this like almost a death of documentation?
07:53Basically. Right? Because code is the best source of truth.
07:58Now there are people that might be like, oh, but isn't this going to inflate the context window? You see, in the old days, the harnesses used to index the entire code base and they would use vector and rag and all that stuff.
08:10Now the models are so good at search. All it needs is a search tool. So you just need to guide it where to search to.
08:18Right? And it's going to reference the code base. It's going to find the exact function.
08:22No guessing. It's gonna find the exact function. And I can't even tell you, like, eight out of 10 times, like, using this, like, I'm getting it spot on.
08:30Yeah. I think the point is that, like, if you do your job properly in the context engineering,
08:34you actually save so much time on testing and debugging because then you don't have any errors. Yeah. Context engineering is like, it is it's so important.
08:43Right? Like, if this is your 227
08:46k context window, your agent is smart probably up to, this level. Maybe I'm being a little mean. Maybe you could go a little further.
08:54But, like, the more you bloat it, the dumber it's going to get. So you wanna stay in a sweet spot like this. Right?
09:00So being able to give it the exact snippet it needs or explaining exactly what you need. And this is why agentic engineering will yield better result than vibe coding because in vibe coding, you're offshoring the thinking to the agent. Right?
09:14Yeah. In agentic engineering, you're doing the thinking and then you're just . You're letting a bunch of junior grads who are very cracked but need a lot of guidance do the work.
09:26So if you have, like, a 272 k context model, I love codex because it'll tell you right here. Like, at 77%, unless I'm done here, this is, like, I just need to start a new thread.
09:38Like, this is too much information for the agent. Now I know a lot of harnesses have, like, slash compact. What's funny is with Cloud Code, David, even, like, one of their I forgot which developer it was, but one of their lead devs in, like, a tutorial he was showing, even he didn't use slash compact, like, the compaction engine that Mhmm.
09:57Cloud Code has. He started a new session. Right?
09:59And that's Yeah. And it's slow also super slow. You know?
10:01Yeah. Right? So you just start a new session, but I say that all to say, this tool allows you to have super precise prompts.
10:09Like, the agent doesn't have to do web search. It literally like, you're giving the best documentation, which is code. So so far, harness cursor with, uh, g p t five five extra high, uh, fast.
10:21We're using, uh, Slopus 4.7 max. And then for tools, open source, which again is completely open source, is a great, like, way to, like, uh, context engineer your prompts, giving you the exact reference it needs.
10:35And a tip for context engineering is you just wanna keep it low. Like, you wanna keep your features minimal.
10:42And that's the way you keep this short is, like, it's being very methodical with, um, like, your feature plan. Like, for example, I don't know if you saw an x, David. A lot of people are debating, um, should the AI do the plan?
10:54Should it not do the plan? Should you jump in straight to the feature? I do the plan, not even for the agent, but for me.
11:01The plan is a great way for me because sometimes I'll run multiple threats. The plan is a great way for me to hold the agent accountable. So you'd be surprised there are lot of times when you create a plan and it sounds like a good plan and then you tell it to generate and it will start to fumble somewhere.
11:17And the reason being is the model doesn't think about its context window when it's generating a plan. You've given it a task. It's going to dump everything in that plan.
11:26It needs to complete the task. So if you have a large task that requires this much context, it's never good to do it right.
11:34But I'll get it to generate a plan and then once the plan is generated, I'll be like, oh, this is a little too big. How can we make this super small PR, super small chunks that are easy for you to review? Then I have like multiple action steps.
11:47Right? I say that all to say, context engineering might as well be a a principle in engineering, uh, in and of itself because this is a a make or break an output of how good things will be. And I guess the principle is, like, staying in charge.
12:00Right? Like, even though the AI is writing basically all of the code now, you still have to stay in charge at some degree because at any point you can say, what are the issues with this code base? And even if there are no issues, it will happily suggest 10 issues.
12:11Right? So, like, people who are new to this, they, you know, relinquish the control way way too easily and trust the AI with too much with the things that they shouldn't trust at all. Yeah.
12:21Like, the model doesn't think the way humans think. Right? Like, I I don't know if anyone who's watching this noticed, but when Chad GBT, when OpenAI shut down four o, there was a huge, like Oh, yeah.
12:32That was huge. People who are protesting. And because they had developed a relationship with the model, the model was so agreeable that people started to fall in love with it.
12:42Right? Crazy. A lot of people might think that's weird.
12:44But when you're building applications using code, that actually happens to a lot of people where they think they're really smart and, like, the model is, like, like, leading the right direction. You're like, you think you've made a right decision, but the model has no idea. It just predicts the next text.
12:58The next text can be wrong. But when you're in the driver's seat guiding and making the decisions, like, you have to almost treat this like a that knows everything but doesn't know how to know use everything.
13:10And that sort of relationship allows you to build again. Three months, four months, 90% of the code, 95% of the code is AI generated in my work.
13:18And I'll get cooked if I draw if I if if bad code gets shipped. I'll continue on the second tool. So the second tool is a a skill of mine.
13:27Now anyone can generate this especially if you if you have, like, development background or you can, like, go back and forth with it. One thing that the agent will do, the agent has a tendency of rewriting code that exists. So let's say you have a function that streams a response from an agent.
13:42Right? Like, you're basically, you're building a chat functionality. You're you're building the response, uh, function where, like, it streams the message once someone has sent a message.
13:51And then you're like, you know what? I want to integrate Telegram. You tell the agent I wanna integrate Telegram and it'll build that for you.
13:57Almost nine times out of 10, it's not reusing the function that's already been created. It's rewriting a new one. And the reason why this becomes a problem, you start to have code smell.
14:05Code base is getting large. There's too many moving parts. It it's going to be hard for the agent to debug what's going wrong cause there's many, many touch points.
14:14Right? What this skill does, it will, like, structure the code in what's called a service layer. And, basically, what a service layer does, it just creates these these functions that can be reused again and again.
14:24What I'll do first is I'll generate the feature. Right? So I'll build the feature first.
14:28Right? GPT five, uh, x high, and then cursor. And then I'll test, obviously, I'll test locally.
14:34Once I like where I'm at, I know when I later on when I come back to this feature, because I'll probably forget how it works. I'll probably forget where the code lives. The the code that was probably written is not the best shaped code.
14:45Now the model is capable of doing this, but out of the box, it won't do it for you. Right? It's like a lazy person.
14:51If you tell me, David, I want you to do this task for me. You know, human nature dictates that I will probably take the least path of resistance. Right?
15:00The model's probably gonna do the same thing. So, basically, this lets anybody talk to the model as if they were like a staff engineer. Basically.
15:07Right? So, like, you can see over here, I have all these, like, plans. By the way, again, everything we talk about in this podcast, all the skills, all the tools of the repos are available below the video.
15:17Click on the first link. It's completely free. So go ahead and grab it.
15:21Now, again, this isn't really for the agent, but more so for me to remember. And look what the goal here it says. It says make deployment provisioning and repair simpler by moving repeated runtime mechanics, watch this, behind deeper structured modules while keeping convicts actions responsible for domain policy.
15:37So there are repeated runtime mechanics. I only need one. I don't need five or six of them.
15:42Right? And this causes issues. I run this skill and I tell it, look at the code base.
15:47Where do you see duplicated code? The h will pop up. It even brings it up here.
15:51It says, modify this file, modify this file, modify this. So these are files where I have repeated code that shouldn't be repeated.
15:59So with this type of skill, and again, free for anyone to to check out Yeah. We got a link at the law. It allows your code to be cleaner.
16:06And what I find is when the structure of the code is cleaner, it's easier for an agent to pick up on a new session. Right? So if it's just messy and it's, like, all over the place, if it's hard for a human to reach, it's probably gonna be hard for an agent.
16:17Definitely. Right? Clean structure matters.
16:19Thank goodness. You don't really exactly need to know a 100% what that looks like. You can use a skill like this.
16:24There's also another person. I think his name is Matt Pocock. Yeah.
16:29He has a very similar skill. I would say his is more for if you're a bit more technical, it helps. If you're not, you might shoot yourself in the foot, but he has this improved code based structure.
16:39This is another helpful one as well. Right? So I say that to say, after every feature that I built, I run one of these because I know push comes to shove.
16:47If the agent needs to work on it again, it's probably written in a way what's gonna confuse itself. It's like applying to, you know, age old principles, but, like, in the new way.
16:55After you build something, you document it for others, but now it's not for the human. It's more for so for agent. And and what's funny is a lot of the old, like, engineering practices that, you know, a lot of engineers shun because it was difficult or made life a living hell is actually great for agents, believe it or not.
17:09So, like, the code being well structured, they're being defined tests is really great for the agents, uh, this being another. So I use open source. I use code structure.
17:18I'm going to recommend a, like, a specific tool. Like, this isn't me promoting a tool, but, like, I just use it and I could prove that I use it. I use greptile for code review.
17:27The reason why I like greptile is for actually a specific skill called greploop, and this skill will make it feel like you're using auto researcher, uh, from Karpathy. I can maybe, like, explain through an example.
17:39I have this PR right here, PR 80. I've done some changes giving the ability for an agent to, like, fully ingest every row of a spreadsheet. Greptile gives me gives me these confidence scores.
17:51Four out of five, three out of five, five out of five. Let's say I get a confidence score of three out of five, and almost always, it will give me like the issue. It'll tell me why it gave me three out of five.
18:01Right? And I'm not going to say this is perfect, but from my testing, it's pretty darn good.
18:06Right? Especially if you're a non technical person, you need this. Like, you you you need need need need this.
18:12So let's say I get a three out of five score. I'll install the skill. All I do is slash grep loop and I hit enter.
18:18The agent is going to read the PR. It's going to read the feedback, and it's going to fix the feedback, and it's going to wait for a new review to happen. So let's say it was a two out of five, and now let's say it gets a three out of five.
18:29You know what it's going to do, David? It's gonna keep going. It will not stop until it gets a five out of five.
18:35This is why I led up with all this first. I talked about context being important and then these tools help you structure the code because by the time you get here, you're going to have a simple feature, a minimal PR, and the code structure's clean. GREBTIL, the review agent reviews it, you're going to get maybe a three out of five to four out of five.
18:54Then what I do is I'll literally run slash and then I work on something else. And by the time it's done, I get almost nine times out of 10, I get a five out of five. The instance I don't get a five out of five is because I gave it a PR that was like 9,000 lines, 10,000 lines, 12,000 lines.
19:10We have to understand they're probably using the same agent, the same model that we're using. So context window is finite, so we have to be smart about it. If I respected the initial rules of keeping context minimal and, like, I ship this PR, like, 60% and then GrebTal is reviewing this now, the errors it gives me are the exact errors that exist in that feature.
19:30GrebLoop is going to catch them and fix them. The last three months, I have, a a pretty big app I'm going to ship soon, the entire thing iAgentic engineered.
19:40That doesn't mean I read the code. Like, I looked at the structure. I have another developer working on it.
19:45He's been asking me questions. You know, knowledge still matters. Even if you don't understand the syntax, which syntax doesn't really matter nowadays, understanding how, like, good code and architecture works helps.
19:56Yeah. Man, I'm telling you, I am going far and fast. The app I'm launching soon, if I told you I built it in two weeks, three weeks, a month, you would not believe.
20:05But with this exact formula, cursor is the harness, the best of class models. If someone says, can I do this with Kimmy? Can I do no?
20:12You have to have the best in class model. Like Yeah. They are not OpenAI or Anthropic.
20:18Yeah. It's true. It's just a night and day difference.
20:21Right? Because I know there's a big local model movement. I love that.
20:24I I have a Gamma running on my machine, but it will not write code like this. And then you give it open source to download whatever packages or libraries you're using, and then you structure after every feature. When you make the PR and you run greploop, notice how it said, oh, PR 80 is good to merge.
20:41Greptile five out of five. Right? So it would have kept on going.
20:45Like and I have times where, like, it went off for, like, twenty, thirty minutes where it's like, ah, the agent I like, cursor will tell me, oh, I made a mistake and I missed this. Greptile caught this. Deploying a fix now.
20:57Pushes it to GitHub. Greptile then sends the review.
21:01When it's five out of five, it stops automatically. Right? And the one thing about the super, um, super duper models, Opus four seven, and particularly GPT five five extra high, they write a lot of tests.
21:14I don't know if you noticed this, David. They write a lot of tests. And this is actually great because now what the model will do is whenever it gets, um, whenever it gets feedback, whenever, like, oh, this is broken, it's going to write the test in which that feature would have worked, that bug would have worked, and it won't stop working until that test passes.
21:31Even though auto research was, uh, interesting and unique paradigm by Karpathy, that's essentially what agentic engineering has been accomplished. You give your agent just enough tools and just enough context for it to work in a loop where by the time it's done, the feature is done. The reason you're using these skills and, you know, providing the right context is that so the end outcome is clear.
21:50Right? Whether it's a test passing,
21:52whether it is the app is deployed, whatever. You know, I guess that's the main idea behind also the slash goal feature inside of Codex is that you give it the end state. Right?
22:00It's not just like a dumb loop that runs forever, which that's how people understand it who don't use it. It's all about how well can you describe the desired outcome. And if you nail that, the agent can basically figure out a way to get that.
22:10I think of it like this. Like, let's say, like, you and me are on a basketball court and you were telling me, Michael, this is how you shoot. Right?
22:16You shoot like this. Like, you've explained it to me once and properly.
22:21I probably won't get it right the first time. I probably won't get it right the second time. You might interject the second time, give me some more feedback.
22:27But then by the third time, I get it. Fourth time, I get it. Fifth time, I get it.
22:31That's basically what we're doing. All this is a giant massaging machine to the agent to you you give it just enough information, and it'll do the right thing.
22:42And that's what this is. Right? This is agentic engineering as a whole.
22:45Giving it enough info, enough guardrails, enough of a feedback loop. Right?
22:50If I told you, you know what? Uh, no open source, no code structure, uh, like, like, thinking about the feature.
22:59I just tell it, yo, build me this and slash greploop, we have Ralph Wiggum. And we've all seen where that led us.
23:05Right? It was cool at first. I even have, like, a custom implementation called Ralphie, but you're not going to get far.
23:11Right now in engineering, human approval, human thought really, really matters.
23:17And I've noticed a lot of people who are building, launching apps and stuff, they're just letting the agent think for them. So it's actually a great time to stand out. It's actually a great time to to build something that feels unique and that actually works.
23:30And this is, like, my process. Another thing I would also say is, like, using popular tools, but not just popular tools, tools that are very codified.
23:40Right? For example, I use a library called Svelte overreact.
23:45And the reason why I use Svelte overreact is even though it's a much newer framework, and a lot of people argue that the model hasn't been trained enough on it, a lot of the syntax is HTML and TypeScript, which the agent is great at.
24:01Yeah. Right? Versus React, especially if you're using a newer version, there's all these hooks.
24:06There's all these foot guns. But with Svelte, it's very, like, it's very to the core principle of HTML and TypeScript.
24:14And here's what's cool. I practice what I preach. I'm pretty sure in this project, if I go to repos, you're gonna see the Svelte JS Nice.
24:23Code base under open source. Right? So if there's something that I feel like it's not doing right, I'll be like, reference the Svelte JS code base.
24:30Right? And it's going to use standard practice, the best practice. So I use Svelte.
24:34So is this headed to, like, people are gonna be building basically only on open source projects, frameworks, and stuff like that because of this reason? Yeah. Like, because code is the best context.
24:45Right? Like, think about it. Like, we give like, people are debating markdown files or HTML files.
24:51Like, it's not straight English. Right? It's some sort of codified structured language.
24:56Right? So code is the best context. And if a company is building a developer tool, uh, again, this is a business decision, they have to open source.
25:06Right? Because when the agent like, human written docs are the worst. The The worst, the worst, the worst.
25:11Now I know there's startups where you give it the repo and it'll generate the docs for you. But again, why am I why are we taking that extra step when you can just give me the code? Exactly.
25:22Right? That's such a good insight.
25:24There's a there's this other library called effect everyone's been raving about. It's like a a TypeScript library. It's super, like, super typed this to stat.
25:34You know what? I'm not going to care. I'm going to find the repo, which it's a library, so it should be open source.
25:41Right? I'm gonna take this, and I'm gonna use open source to download in my code base.
25:46And now me and my agent are gonna go back and forth on me setting it up. So even with frameworks I'm using, uh, in this case, I'm using Svelte, I'll use open source.
25:55Um, for back ends, again, I know people are fans of Supabase, and if it works for you, you can. And I know I work at Convex, so but in all honesty, I use Convex, and here's the reason why. Everything in Convex is is code.
26:06The only time you have to go to your dashboard is when you wanna set up a production instance or your app is blown up and it's time to pay. I wanna set up scheduled functions, meaning I wanna close my app and have something run-in the background, which your app should have that. Right?
26:21It's code. I wanna write a API. It's code.
26:25Everything convex. All the features is TypeScript code.
26:29Now why is this great, David? The reason why this is great is because the agent has full context on what my back end is doing. It's not it's not guessing about the schema and you know?
26:38I don't need to take a screenshot of the dashboard and tell it this is what's happening, and then tell me I don't know if anyone's ever done this. Tell me what tab to click sometimes. You get a little lazy.
26:48But Convix is all code. Right? And if you notice everything I'm saying, it goes back to really giving the agent the perfect context it needs.
26:57Right? So the back end, it's code. The framework, I give it access to the code base.
27:02When I combine these two with this structure, I've been able to ship, like, almost anything and everything.
27:10Now it's not a smooth sailing. Like, obviously, when you build anything, and I think this is a mindset I know you preach about a lot, that's lost.
27:17People understand that they have to work hard either in the gym or in life. But for some reason, when it comes to agentic engineer using an agent, if it doesn't get it in one shot, it's over. Yeah.
27:26Right? Now you have to have some sort of audacity.
27:29Right? Like, you're building software that okay.
27:31If it's an internal tool, fine. But if you're building software that you want people to use, like, care is important, being thoughtful is important, spending time is important.
27:40Right? You're gonna handle people's credentials and information. Us to learn.
27:44Also, like, people who don't come from developer by account, like, close their mind to any technical concepts. Right? It's crazy.
27:49Yeah. Yeah. Like, someone will be like, oh, it's like a CLI terminal thing.
27:52It's like it's actually so simple. Like, it's actually so simple. And what's funny is I'm starting to build on technologies I've never really messed with.
28:00Like, I'm not a big network engineer. And I've been using, like, the agents to, like, explain to me and, like, clean up my desktop and do these things. Just yesterday, I don't know if you saw it on Twitter, some guy who lost his wallet password Oh, I've seen it.
28:14I've seen it. He used ClarCode. Right?
28:16Crazy. So if anything, these tools should expand your mind.
28:21You should be delusional and crazy. You should think you can build anything, and maybe you can't, but dang it.
28:27At least try. Right? Like, at least give it a go.
28:30Give it an honest effort. You know, all the motivational David David Goggins stuff, you know, fail forward, yada yada yada. That same stuff applies to engineering.
28:38But I find with, like, these tools and this combination, I've been able to get a really, really solid output.
28:47Yeah. And it's crazy to think that, like, most of, you know, vibe coders, they don't have any of this. So maybe they only have one of these tools, but they're completely unaware of the other principles and the other tools that exist, or maybe they're using a cheap model or a free version of some tool.
28:59You know? You meet it every day. I meet people every day who think, like, they're on the cutting edge of AI, and they have, like, the free SharedGBD or a free cloth.
29:06Right?
29:07Yeah. Like, unfortunately, this is going to be a a, like, a money game.
29:14Right? You know, we're being subsidized now. At some point, the subsidies will end.
29:19Right? And it's going to be, like, the people who have money to play, like, you're going to get better results. So I think, like, even though, like, again, $200 a month is a lot, and I understand everyone's in different situations.
29:32Like, if you're a young person, right, I'm not gonna speak to people who have kids and responsibilities and mortgage because that's a whole different, you know, life. But if you're, like, young and you have a job and the reason why you're not paying the 200 subscription on Codex is because you're drinking out with the boys
29:49Crazy.
29:50You know what I mean? Like, again, for the to the adult who has Even the 100.
29:54Like, even even the 100 Kodak subscription gives you so much, like, OpenAI subsidizes it like crazy.
29:58They are. And, like, to be honest, like, as long as like, if you want to continue the subsidies to go, go on Twitter and glaze OpenAI. They'll probably keep pushing for it.
30:07Right? So, like, getting, like, a $100 sub with Codex, in my opinion, and just pushing it to the limits, and now they're, like, turning it into, like, a super app where, like, you could do stuff and, like, workflows and stuff like that, I definitely think it's worth it.
30:23I think it's worth the investment.
30:26And a lot of these way, not even coding. Right? Like, imagine what that gives you.
30:29People my only see the coding is like, don't know if that's gonna you know, I'm not a developer. I'm not building that many tools. But, like, the difference between paying $100 a month for shared GPT versus free plan is insane.
30:40You get 5.5 Pro extended, and you can, like, literally have, like, a professional lawyer, a professional doctor at your fingertips. And, like, anything you're dealing with, you know, maybe, like, some some dispute with somebody on your team or somebody's threatening lawsuit, could just launch a deep research. And within fifteen minutes, you have, like, a insane response
30:56that, like, would cost you thousands of dollars if you paid human lawyers. I'll I'll I'll give you an example. Like, real life example.
31:02I was sent a contract for some work that I that, you know, a company wanted me to do. And when I tell you that contract was, like, 27 pages. Now for a long time, I was a boomer, and I had a lawyer for contracts because I was like, oh, maybe, like, you know, an agent might miss something.
31:17But the lawyers are expensive. They they they charge, especially in Canada, America. They're very, very expensive.
31:23Yeah. So I was like, oh, let me let me give it to Claude. Let me give it to Claude desktop.
31:27When I tell you every page highlighted with every single point, giving me a rebuttal, I ended up getting, like, the money that they were going to pay me three x because of the nonsense I saw in the in the in the in the contract.
31:43And then I gave Claude, like, the analytics of all the stuff that I've done, the work that I've done. And the agent was like, yo. You need to charge more.
31:50Like and I literally was telling the agent, oh, but I'm scared. Like, I don't wanna lose. Like, I literally was telling was like, I'm I'm scared.
31:56I'm not a salesperson. You know? I don't wanna lose the contract.
31:59And the agent was like, no. Don't be like, it literally was like a coworker. And then it wrote a message.
32:05I then rewrote it in my language because, you know, you don't wanna you don't wanna send AI messages. Yeah. Of course.
32:10And then they responded back saying, yeah. We'll change the terms, and we'll three x the price. If if that was $200 a month, that was infinitely times way more important.
32:20Right? So
32:22I think trying is the best thing. Like, trying is example. From like, a similar thing where it saved me a shit ton of money, it's accounting.
32:30Right? So I had to, like, redo accounting for 2024, twenty twenty five five for one of my companies.
32:36I got accountants, and, again, not some crazy expensive firm, just like normal accountants. They quote or they said, like, oh, this is, like, 3,000 transactions. This is gonna be 42,080, which is the emiratic currency.
32:48Let's say, like, 5 or $6,000. Right? I'm like, wait a minute.
32:51Yeah. You're charging me based on transaction amount? And I was like, listen.
32:55Let's check the accounting software. Do they have API? They have API.
32:58So I'm like, no. I'm doing this with Kirsten and Clothecode. I sat down for, like, you know, two hours, and it's done.
33:04And I saved literally 5 or $6,000, And I honestly trust Kirsten Klotzkor more than some, like, you know, accounting associate at that firm.
33:13Yeah.
33:14Who who like like, again and it's it's these things that even though we're talking about, like, agentic engineering and development, I'm actually more bullish on knowledge work.
33:24Like, there's so much mundane. Like, especially if you're running a business or you're an entrepreneur or, you know, you have a side hustle or you work a corporate job.
33:33Like, I have friends who are really hub David in a corporate job. And when I tell you the amount of productivity in generating reports, spreadsheets, all this type of stuff that people would spend hours on that, like, a $100 subscription saves you on, like, if you haven't made the decision again, I don't get paid by any of these companies.
33:50I wish I did. I wish I got some stock.
33:52But I'm personally telling you because, like Bless us. We have unlimited.
33:56But, like, I'm genuinely saying, like, if they bumped it up to $500 a month, I probably would still pay. I'm paying. I'm buying it.
34:02Yeah. You see? Okay.
34:04So Dave's not even thinking about it. So you see, like, getting exposed to the tools and using it is probably the best thing.
34:11And I don't wanna sound like a super, like, you know, like, one of those billionaires on TBPN, but I generally think, like, you have an advantage when you use the new tools. Like, maybe there's something better that comes out next year, but the fact that you have exposure with the tools now means when the better tool comes next year, you'll be better equipped in using it and learning it, and you move much faster than someone doing it for the first time.
34:32Absolutely. Now one thing I wanna touch on because, you know, you mentioned you you've been building something for, like, three weeks. A lot of people get stuck in the building phase forever.
34:41Right? I I know many people who are, always when I follow-up with them, they're like, yo. I'm, like, two weeks away from launching.
34:46I follow-up in six months, and they're two weeks away from launching. What advice would you give to these people? Yeah.
34:52So I was in San Francisco recently, and to the people who experienced this, I was experiencing this too until I went to San Francisco.
34:59I don't know, David, if you've been there. The level of delusion, and I'm not saying in a bad way, but the level at which people believe they will succeed is so high.
35:09Like, I know you have to, like, believe in yourself. Me, most people do not believe in themselves the way the people in San Francisco do. That's number one.
35:17They have this level of belief that whatever they're building is gonna change the world. Like, someone will tell you, I'm building an AI UGC app. Right?
35:25And there's a couple of them out there. I'm sure they can, like, you know, make it better in some different areas. Let's say AI UGC app.
35:31When they explain it to you though, you feel like this is the greatest app ever. And at first, I thought I was like, oh, maybe people are just grifters. No.
35:38No. No. No.
35:39They genuinely believe it. And because they believe it, the moment they have a semi functional MVP, it might not even work. Auth might break when a 100 people get on the site.
35:49They'll launch it anyway. They'll launch it. They'll get hype.
35:52And guess what? They'll raise the money. And then they'll hire people or they'll spawn a bunch they'll do whatever.
35:58And at some point, their product gets good. Meanwhile, you and me are .
36:04We haven't launched yet. Those guys have raised $10,000,000. Right?
36:06So I say this all to say, whether you're bootstrap raise, whatever it is, you you you gotta launch early. And I know it's hard because you don't want people to catch the bugs that you catch.
36:17But believe it or not, a lot of people are actually most people are actually invested in cool new tools. I mean, there's communities.
36:26Forget technical people. There's communities who search for new Mac apps to try out.
36:31Right? So Yeah. People are you know, most people are bored, and they wanna try out something new.
36:36Right? And they're willing to try your thing out. And if it sucks, they'll give you feedback.
36:40You fix and you go again. And your product is better than just like keeping it for yourself and like you to lose yourself, you know, that it's good when you have no touch with reality. And I just want you and this is what I want people to want you to go on Twitter and I just want you to look at all those launch videos, all those animated videos.
36:55You know why they're animated? Because the product barely works. Yet they're launching, they're getting more users, they got more MRR than me and you, and what are we doing?
37:04We're just fixing this one more feature. It is it's the biggest scam. I'm fighting it myself, but it literally is the biggest scam.
37:13And time after time, I see it all the time. You have to be delusional. Right?
37:20Again, I'm not saying be dumb as in, like, you know, don't make dumb decisions. Don't say, you know, like, for example, like, Dell, don't say you are the best SOC two platform and you're actually committing crimes.
37:32Don't do that. But, like, just believe in what you're building. Believe in the abilities that you present.
37:38You know, communicate a vision that, like, okay. It's it's this right now, but it will be this in the future, and share with the world.
37:44Build, like, you know, in public. Not the community, but, like, build in public. And don't be in the shadows because there are gonna be many competitors.
37:52Right? Unless you're, like, Andrew unless you're Karpathy and you're building some sort of new paradigm. If you are, then stay in the shadows.
37:59But if you're not, you have competition, and they're gonna move fast, and they're gonna burn more tokens than you.
38:05And their app might not even be as good, but they're gonna win. So it might as well be us.
38:11Um, that's what I, like, that's what I literally told, like, a couple people, like, last couple weeks who had really great apps and but they're like, ah, just one more feature. And I'm like, yeah. Your competition will smoke you if you keep going like that.
38:23Yeah. So, like, the delusion and the belief is, like, half of the success. So many other things fall into place.
38:28And also, if you believe in it so much, like, any roadblock, you just, like, brush it aside. Right? But if you're kinda, like, on the edge, then, like, you quit.
38:36Yeah. Like, I I'll share maybe, like, a personal example. Like, I have a full time job.
38:41And after the full time job, I'm working on this app that I'm watching very soon. And, like, the cost of that is sleep. The cost of that maybe, you know, get togethers.
38:50Right? So Social events, yeah, fun, whatever. Whatever it is.
38:54Right? So, like, if it's like a product that you're really bullish on that you see yourself using, then, you know, just launch, get feedback.
39:02If people dunk on you, great. Now you release v two and, you know, be funny with your marketing. Again, I'm I'm not the best marketer, but, like, go being out there and and and and building in front of people, people actually like that.
39:16People love the underdog. No one usually gets hate as an underdog. You get hate once you've succeeded.
39:22So might as well absorb all the love now and succeed, and then later on, you can figure out the rest. But, yeah, please launch quick.
39:30I wanna talk about how you think about security in the ancient era. Right?
39:35Cybersecurity. We see, like, every day is a new breach, new vulnerability, new hack, new attack.
39:40Tell tell people who are not as technical as you, how should they think about this? Yeah. It's cooked.
39:46We're cooked. I'm a just keep it a book. It's it's it's really scary.
39:50Right? It's like, I would say maybe don't be super famous.
39:56Like, don't antagonize a specific group of people who wield power or, like, technological like, all the jokes aside, it really is, um, a scary time because the models are starting to get really good.
40:10And you can imagine if you can run this in a loop to fix a feature, you can run the agent in a loop to do very nefarious things. Right? And even now, you go on Hugging Face.
40:20There's a lot of distilled models that the guardrails are removed. Like, you can prompt it something nefarious, and it'll do it for you.
40:29For for people personally, and this might be a little crazy. Like, for example, in my family, we have a passphrase.
40:36Right? So if you get a voice that sounds like me asking me for money, you're gonna ask what the passphrase is.
40:41Right? Because the voice cloning has gotten really good. Two f a is mandatory now.
40:45It's not even
40:47it's not even a Password managers. It's crazy how many people don't use password managers.
40:51Yeah. Please. Like, one password something.
40:54Right? Like, you need to have super complex passwords and then the key that one password gives you, take half, give it to, like, your mom, your dad, or someone you trust. Like, we have to be very, very secure with, like, the accounts and the stuff that we have.
41:09Two f a, not via text. I got sim swap not too long ago. It's happened.
41:14It's real. If people wanna get you through sim swap, they will. You use a Authenticator.
41:192 f a, like, Google auth, whatever. Right? And but most importantly is, like, if you're in the agentic space, especially, you're building tools and stuff, I would be very careful with the packages you download and stuff like that.
41:33Right? Because for the most part, I would say the the consumer apps that everybody's using are on very, very high alert, and they've been getting hacked before.
41:43They'll probably get hacked again. The dangerous attack vector is when you're downloading packages, um, and code.
41:49Right? And something you can tell your agent, you can really prompt your a you can just prompt your agent this. You could tell your agent, never install a package that is that is let that is younger than fourteen days.
42:02Right? Because a lot of the attack vectors are packages that dropped literally in the last couple days or last couple of hours. Um, that's number one.
42:10Right? So telling your agent to never download a package that's, like, fourteen days. Like, that's not older than fourteen days.
42:17Number two is generally being part of the discourse on Twitter. Right? I think Twitter is just a great place.
42:22If you follow the people David follows or I follow, you'll probably have your algorithm seasoned enough where, like, when something happens, you're one of the first people to know. And third, this is actually going to probably be a sector in the space that's going to keep on growing.
42:37Right? So if any of you young people watching this are in school for security or any such thing like that, keep going.
42:44The job market's looking really, really good for you. But Yeah. All in all, from consumer apps, two FA is a must.
42:52Also, tangent, if you have any older people in your family, please explain to them these things. Right? Because I've heard stories of people's, you know, grandfathers sending money to strangers thinking it's like a a granddaughter or an old lost family member.
43:08Right? Like, for us, it might not make sense, but for an old person, seed dance is convincing enough.
43:14GPT Image two Oh, it's I mean, GPT Image two is near perfect. You know? Yeah.
43:18So, like, definitely, like, the old people in your family, like, get them to be like, get them to ask you first. Like, the con of this is, like, even on my WhatsApp, I have, like, six different messages from older people in my family saying, oh, is this AI?
43:31Is this a scam? And because, like, they're easy targets, this is something I'd recommend, like, people communicate with.
43:38But all in all, it's two FA to, you know, one password have a password manager,
43:44um, really, like, don't I would say one more thing on on what you added with the Twitter. Right? When when you spot something within minutes, you can just have, like, chat gbt, perplexity, research it, get, like, every information about this breach, and then tell it to give you a one paragraph instruction on how to analyze your MacBook to see if you're exposed.
44:02You paste that into Clothes code, pi, Hermes, whatever you're running, and it will tell you, like, yes, this package is in this project or no, you're safe. And, you know, that's probably the best thing. Yeah.
44:11And that's a big thing. Like, especially, like, Claude is really good at this. Like, you can, like, literally copy the tweet, paste it, and say, like, am I cooked?
44:18It'll understand.
44:20And it will, like, read the files. I think there was one in particular. I forgot which one.
44:24There was one in particular where I thought I was caught, like, because I download the package literally, like, a version before, and the next version was a scam. And, literally, Claude looked through every single directly, looked at the system directly, and it checked.
44:38And it's like, oh, your machine is clean. Right? So, uh, doing a lot of that is necessary, but also, again, being smart with what you install.
44:45Right? That's how the big attack vectors are happening now. It's through packages.
44:49And if you're watching this video, you're probably building with agents. If you set up that prompt where you don't install a package that's younger than fourteen days, by then, the vulnerability would have been caught and you would have been safe. So, like, that's the biggest thing.
45:02And on the consumer side, be smart.
45:05Where do you see the future headed? Like, if you had to say, like, in three months or six months, what's gonna be possible?
45:11Yeah. It's like, I'm more bullish on knowledge work than I am on, like, agentic engineering just because I find that there's too large of a surface area in engineering for the models to just be good at all of them.
45:27Like, if you notice, g p t five five is a tad bit smarter when it comes to things of architecture and, like, back end stuff.
45:37And then Opus really seems to be great at UI stuff. Right? So there's, like, these two large of a verticals.
45:42And now the question is, does that mean we get we make a bigger model, or do we make smaller specialized models? Right? So there's a lot of, like, thinking there, obviously, the model providers, you know, have the smartest people, uh, figuring that stuff out.
45:56But I'm more bullish on knowledge work because I think the issue now with a lot of knowledge work is tools not being built around it. I actually don't think it's not the models being not capable enough.
46:07I think the models aren't there for knowledge work. I think we have smart enough, models where a ton of knowledge work can be done through them.
46:13We just don't have the tooling around them, which is why you have both OpenAI and Anthropic launching these, like, consulting companies. Right?
46:21And they're gonna deploy forward, deploy engineers into companies and small businesses to help them set up. So if you're in a place where that doesn't use a lot of tools and you wanna get a little promotion, maybe sharing these things.
46:33Right? Like, using these things, being an example of these things. Be a leader in these things.
46:37Like, I heard David, some guy that literally I I I just met in an event, and he was telling me, like, no joke. Like, he at his job, he helped them.
46:46Like, they do they they they do, like, a lot of, like, contracts. Right? And they're not a law firm.
46:51They're not big enough to afford, like, a, like, a proper law firm. So, like, they pay these guys on retainers. It's a lot of money for them.
46:57He basically did one presentation in front of them, showed them how Claude Coburg worked, and, like, they made him a manager. Guy's, like, 24. Right?
47:03So, um, because, like, again, he showed them with knowledge or how, uh, valuable things could be. So I think in the next three to six months, we're going to start seeing even a boom in knowledge work.
47:13In terms of agentic engineering, I'm really hoping Opus five comes out, and it just blows us out the water. Because if everyone remembers December is what changed the game.
47:23That's when we realize, oh, this is a shift. And I I think that's what Opus is being that's what Anthropic is gearing up for.
47:31OpenAPI is taking too much share. I know they're gonna stop it.
47:35I hope they are because this competition helps. So in three months, knowledge work is gonna continue to boom. Maybe we get a new model.
47:42But in terms of the long term effects of all this, I actually have no idea. And that's kind of terrifying because, like, you know, when you when people say, oh, jobs are gonna get replaced this this step.
47:54Not fully, but some will. Yeah.
47:57Right? And what does that look like?
48:00Right? Like, what does, you know, being job proof look like?
48:04What jobs are safe? Right? I I mean, if you told me three years ago, the agents would be great at coding, I'd laugh at you.
48:09I'd tell you programmers are gonna rule the world if anything, because no one can do programming, and yet look where we're at now. Right?
48:16So I think being informed, being proactive, and just making this fun, a lot of people are dreading this.
48:22I don't know if you've noticed this, David. A lot of people are dreading, like, the growth in these tools. I think if you just have a simple mind shift shift shift and this stuff becomes fun for you, you actually end up growing with the industry.
48:34Yes. Right? You end up learning new things.
48:36You end up maybe picking up a new gig, a new job, consult whatever it is. Right? So, um, don't take the change as this is happening against me.
48:44But if we have a little mindset shift and say this is happening for me and you're learning with the tools, I genuinely think it'll be a fun time. And whether we go boom or bust,
48:53I'm going to win either way. And I think that, like, that mind shifts of mindset shift helps a lot. Yeah.
48:59I'm I'm glad you mentioned this because, you know, people who aren't as technical as you or on the cutting edge as you, they feel the same way, And they think like us, we are in a club or something, and we we feel we have all the answers. Right? We know exactly what's gonna happen, but the answer is nobody knows.
49:13Right? Nobody knows. It's too much change, too much unpredictability, and you almost have to embrace that.
49:18And like you said, have the right mindset to use that for your own advantage and see, like, okay. If everything changes every six months, that means there is new opportunities for new businesses every six months. Yep.
49:28100%.
49:29And, like, even, like, those of us in this bubble, even our thinking changes, like, every six months. Like, I remember last year when windsurf and stuff were pop popular, the goal was to give the agent all the context.
49:42If you remember, there was even services that would minify your code base into, like, XML, and you would give that as context. Believe stuffing the agent with context was a smart thing.
49:52Now we're doing the complete opposite. Right? So even, like, if you feel like, uh, things have progressed so far, catching up and being ahead is so easy.
50:01And for I'll draw one more graph just for the nontechnical person because I know there's one person saying, oh, but you guys understand this than the third. If this is the if this is a line of you have nontechnical people here and you have super techie people here.
50:18I would say the people on Twitter who really know this stuff that are talking about this every single day on the cutting edge, I would say the peep like, that's it's this it's this section of the graph.
50:30I say this to say that even being right here means you're ahead of most people. Right?
50:36So this idea of, like, I'm nontechnical. I'm sorry. I understand you're not.
50:41You have two options. Either go be technical so you stop complaining or or just have have that dog in you.
50:48You know what I mean? Like, use the tools. You don't know how to use a CLI?
50:51Ask. I don't know how to write Rust, and I properly probably will never learn how to write Rust. You know what I do?
50:57I ask the agent. Okay. I wanna write Rust.
50:59What should we do? You know, I know this is a lost art. Books, maybe pick up a couple books.
51:04Read on some things. Right? You wanna learn engineering principles?
51:08There's tons of engineering books that you can read. Right? There's just to me, like, I'm nontechnical, pisses me off.
51:14Because I have nontechnical friends who are so like, they've who are so on the, like, as you could say, the cutting edge that it's just a matter of grit mindset and maybe a 200 a month subscription. Right? So there really is no excuse.
51:28You just need to be motivated and excited, uh, to use these things, and I think that will take you a a long way. If you're having fun, it doesn't feel like work.
51:36I'm sure if I asked David, David, does the stuff you do feel like work?
51:40The building of AI? No. Like, we're doing that twelve hours a day.
51:43Everything else feels like work. You know? Yeah.
51:45And and, honestly, like, they almost feel like slot machines. It almost feels like a casino because, like Oh, yeah. I noticed
51:51and my wife told me this. Like, every now and then, like, I'll game I used to game a little bit just to, like, you know, relax whatever. I stopped doing that now.
51:59My gaming has been just spitting up an agent and, like, just building some random thing. Right? So, like, I think there's a place where people can make this fun,
52:10and you will find that you'll be a lot more useful in the next couple months when you do that. Yeah. And to, you know, double down on your point, the fact that, like, you cannot say I'm not technical because everything will be technology.
52:21AI will be everywhere. Right? If we see it first in the realm of software, but, like, it's gonna be in the physical atom world soon enough with humanoid robots, drones, and everything.
52:30So the the kind of mindset, like, I'm not technical, it's saying, like, I'm the past. It's like, I'm not future. You know?
52:36If you're new to yourself, like, I'm not future, it's like, woah. Obviously, everybody wants to be future proof and, you know, future ready. So, yeah, everybody will be technical.
52:44The question is how fast you embrace this and how fast you try to get on the cutting edge. 100%. Right?
52:49It it really it it really is a do you have that dog in you question. Right? And I know people who are very much less technical
52:57than me who I would say are on par, maybe even better when it comes to using these tools and these harnesses and configuring them and stuff like that, and they don't even know how to write code. Right? So, you know, take advantage.
53:09Yeah. Alright. So where should people find you, Mickey?
53:12I'm on YouTube. X Ross Mike, r e s m I c.
53:17I talk about cool stuff. I don't post as much as David, but I'm trying. And, yeah, I appreciate you for having me on, man.
53:23Thanks. It means a lot. No worries.
53:24I'm gonna link all your socials below.
53:27Have a great day, Mickey. Alright. Have a good one.
53:29Bye bye. Thank you for watching the entire thing, but watching without action is pointless. So go ahead and apply all the things that Mickey talked about by grabbing the bundle of the skills, the presets, the repos, everything we mentioned in this whole podcast, it's the first link below the video.
53:44Go grab it now and implement it into your own agentic engineering workflow. Again, it's completely free, so go get it now.
The Hook
The bait, then the rug-pull.
Micky Shimeles builds with AI writing 95% of his code and he is an actual engineer. This conversation is the blueprint: the exact harness, the context tools, the feedback loops, and the mindset that separates builders who ship from builders who are still fixing one more feature.
