25:35Jack Roberts · Tutorial
Every Level of Hermes Agent Explained
A 25-minute level-by-level breakdown of Hermes agent, from one-shot prompts to a model-agnostic agentic OS that ships work while you sleep.
June 17thThe argument in one line.
UltraCode replaces the managing LLM with deterministic code so parallel agents run to a guaranteed output shape, eliminating the hallucination and confident incorrectness that degrade single-agent sessions on complex tasks.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code for serious builds and keep re-running sessions because the orchestrating agent forgets context or drifts off-task.
- You want a practical mental model for when multi-agent parallelization saves time versus when it burns tokens for no gain.
- You are building an agentic operating system or personal AI dashboard and want to see how to add new capability modules.
- You are spending heavily on Claude tokens and want a framework for deciding when heavy-compute modes earn their cost.
SKIP IF…
- You are a complete beginner to Claude Code -- the six-mode mental model requires baseline familiarity with the tool.
- You want a deep technical implementation guide; this is conceptual with live demos, not a source code walkthrough.
The full version, fast.
UltraCode is Claude Code's multi-agent fan-out mode, activated by typing 'ultracode' in the prompt. Unlike sub-agents orchestrated by an LLM that can forget or hallucinate, UltraCode uses deterministic code to spawn and manage parallel agents, guaranteeing the task runs to a defined completion state. The video demonstrates two patterns: a judge panel (multiple agents advocating and critiquing competing options with a head judge synthesizing) and a live model-picker dashboard build. The key constraint is token cost: parallel fan-out burns 4-7x more tokens than a normal session, so UltraCode earns its keep on research, multi-perspective audits, and separable build tasks -- not single-file refactors, bug fixes, or step-by-step pipelines where parallelism provides no speed advantage.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:50
01 · UltraCode Just Dropped
Hook and announcement: UltraCode creates up to 10 parallel agents, but only matters if used correctly.
00:50 – 01:15
02 · One Boss, Ten Specialists
Mental model: one deterministic boss script manages many specialist agents on a factory floor.
01:15 – 01:55
03 · How to Activate UltraCode
The effort slider has nothing to do with UltraCode. Activate with keyword ultracode in prompt, or ultracode on for session-wide.
01:55 – 02:47
04 · Six Modes in Claude Code
The six modes: default chat, plan mode, UltraThink, sub-agents, fast mode, UltraCode. UltraCode replaces the managing LLM with deterministic code.
02:47 – 03:29
05 · Judge Panels and Loops
Pattern playbook: adversarial verify, judge panel, pipeline, loop until dry. Each has a defined use case and termination condition.
03:29 – 05:02
06 · LinkedIn Growth Debate (Live Demo)
Live judge-panel demo: three LinkedIn growth options get advocates, skeptics, and a four-person judge panel with distinct lenses running in parallel.
05:02 – 06:10
07 · Why Agents Critique Each Other
Agents hallucinate and go confidently incorrect. Multiple agents ruthlessly critiquing each other in a loop converges on better answers.
06:10 – 06:55
08 · The Full Claude Masterclass (CTA)
Mid-roll course pitch: Claude Code masterclass covering foundation, memory systems, Hermes agent.
06:55 – 08:25
09 · Watching the Verdict Build
Watching 11 parallel agents run live; commentary on token visibility, parallelization speed gains, and agentic OS context.
08:25 – 09:29
10 · Best Model for the Job
Model lock-in is the enemy of high performance. Route by task: Gemini for multimedia, GPT for code review, Claude for general work.
09:29 – 11:40
11 · Context Before You Fire
Pre-UltraCode clarification conversation: get alignment before spinning up agents because UltraCode is only as good as the brief you give it.
11:40 – 13:35
12 · The Finished Model Picker (Live Build)
UltraCode builds a model-picker dashboard into the agentic OS: filterable by smartest/cheapest/fastest/most-used, pulling live data from OpenRouter.
13:35 – 15:26
13 · Locking Down Compliance (Vanta Sponsor)
Vanta for security compliance (SOC2, ISO 27001, GDPR) -- positioned via the host's own Glider startup as a credibility anchor.
15:26 – 17:19
14 · Surgical Versus Max Mode
Surgical (single keyword per request) vs. always-on. The cost warning: 4-7x token burn, realistic usage ~20% of sessions.
17:19 – 19:05
15 · When to Use UltraCode
Decision framework: use for research, audits, multi-perspective reviews, adversarial verification, separable systems. Avoid for single-file refactors, bug fixes, step-by-step dependencies, quick tasks.
19:05 – 19:37
16 · What's Coming Next
Outro pointing to a follow-up video on Claude Code superpowers.
Lines worth screenshotting.
- UltraCode replaces the orchestrating LLM with deterministic code, so the task cannot end until it meets the defined output shape -- this is why it outperforms sub-agents on complex work.
- Activating UltraCode has nothing to do with the effort slider (low/medium/high/max) -- you trigger it by typing ultracode in the prompt, not by cranking the reasoning dial.
- Parallel agents compress long tasks only when the subtasks are genuinely independent; a sequential A-to-B-to-C pipeline gets zero speed benefit from parallelization.
- Agents running alone can become confidently incorrect -- they produce wrong answers with high confidence and no internal signal of error; multi-agent critique loops reduce this failure mode.
- UltraCode burns 4-7x more tokens than a normal session due to parallel fan-out; using it as a default mode drains API credits on orchestration overhead that adds no value for simple tasks.
- The judge-panel pattern works because each agent gets a specific role and lens (CFO, brand purist, growth strategist) -- adding more agents doing the same thing produces redundancy, not better answers.
- Model agnosticism -- routing multimedia to Gemini, code review to GPT, and general work to Claude -- outperforms single-model lock-in as a performance strategy.
- Having a clarifying conversation before triggering UltraCode is not optional; 11 agents will faithfully execute a vague brief in parallel, compounding ambiguity rather than resolving it.
- UltraCode earns its keep on roughly 20% of Claude Code sessions; the other 80% is handled more efficiently by normal mode.
- The completion guarantee is the structural advantage: code-orchestrated fan-out will not exit until the defined output shape is produced, unlike an LLM orchestrator that may declare done prematurely.
UltraCode earns its keep on 20% of tasks -- pick them right.
WHEN TO REACH FOR IT
The token cost is 4-7x normal, so the question is not whether UltraCode is powerful -- it is whether the task genuinely benefits from a team of agents critiquing each other.
04Six Modes in Claude Code
- UltraCode is one of six distinct modes in Claude Code; confusing it with high effort or sub-agents leads to misapplying it on tasks where normal mode would be faster.
05Judge Panels and Loops
- The judge-panel pattern assigns advocates, skeptics, and a synthesis judge -- the adversarial structure is what produces a defensible recommendation, not just adding more agents.
- Adversarial verify and loop-until-dry are separate patterns with different termination conditions; picking the wrong one wastes the entire token budget.
07Why Agents Critique Each Other
- Agents running alone go confidently incorrect without any internal signal of error -- multi-agent critique loops are a reliability mechanism, not just a performance one.
10Best Model for the Job
- Routing by task type requires knowing current model benchmarks; a model-picker dashboard that refreshes weekly solves this without manual research each time.
11Context Before You Fire
- Pre-UltraCode clarification is not overhead -- it is what prevents 11 agents from faithfully executing the wrong brief in parallel and producing a polished wrong answer.
14Surgical Versus Max Mode
- Surgical UltraCode (one keyword per request) is the right default; always-on mode is a deliberate commitment for intensive multi-perspective work only.
- Most Claude Code sessions do not need UltraCode -- treating it as a default burns credits on orchestration overhead that adds no value for simple tasks.
15When to Use UltraCode
- Use UltraCode for: research, audits, multi-perspective reviews, adversarial verification, separable systems where you do not know the answer shape.
- Avoid UltraCode for: single-file refactors, normal feature builds, bug fixes, anything with strict step-by-step dependencies, and quick sub-5-minute tasks.
Terms worth knowing.
- UltraCode
- A Claude Code mode where a deterministic script (not an LLM) orchestrates parallel agents, guaranteeing the task runs to a defined completion state. Activated by the keyword ultracode in the prompt.
- Deterministic fan-out
- The mechanism UltraCode uses: a code-based script spawns N agents and aggregates their outputs according to fixed rules, with no LLM in the orchestrator seat that could forget or deviate.
- Judge panel
- A UltraCode pattern where agents each advocate for a different option, skeptic agents critique each, and a head judge synthesizes a winner -- modeled on adversarial legal reasoning.
- Adversarial verify
- A UltraCode pattern that spawns independent skeptic agents each tasked with refuting a claim; findings survive only if they pass a majority-vote plausibility threshold.
- Loop until dry
- A UltraCode pattern that runs agents in a critique loop, continuing until a convergence condition is met rather than stopping after a fixed number of rounds.
- Confidently incorrect
- The failure mode where a single LLM agent produces a wrong answer with high certainty and no internal signal that it has gone wrong -- the key motivation for multi-agent critique loops.
- Effort level
- The separate reasoning-depth dial in Claude Code (low/medium/high/extra high/max) that controls how hard any one agent thinks -- orthogonal to UltraCode, which is a structural pattern, not a thinking-depth setting.
- Parallelization
- Running multiple subtasks simultaneously rather than sequentially. Only accelerates work when the subtasks are genuinely independent; sequential dependencies eliminate the speed benefit entirely.
- Model agnosticism
- The practice of selecting the best-performing model per task type rather than defaulting to one model for everything, treating vendor lock-in as a performance penalty.
- Agentic OS
- A personal dashboard application that centralizes AI spend tracking, memory systems, model selection, and integrations -- used in this video as the live build target for the model-picker module.
Things they pointed at.
06:55productClaude Code Masterclass (Jack Roberts)
Lines you could clip.
02:01
“Not a bigger brain. A better system.”
05:25
“They can become confidently incorrect.”
08:57
“Lock-in is the enemy of high performance.”
17:00
“It is like firing a bazooka when a hammer very easily would have done the job.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
00:00Claude just released UltraCode, and it is incredibly powerful, but only if you know how to use it correctly.
00:06UltraCode increases the speed, performance, and accuracy by creating up to 10 parallel agents.
00:14This unlocks new capabilities and is an increase on whatever system you were using previously, and nobody's talking about exactly why. And in this video, I'm gonna show you exactly how to use Ultracode correctly and how to build your own agentic operating system.
00:28And the big watch out that you need to make sure you know so you don't make any mistakes, all meaning you can make more money, save time, and get light years ahead of everybody else. And if you're new, I'm Jack. I built a small molytic startup with a gazillion customers.
00:41Now I build my own AI companies, and I share the stuff that works on this channel. So if you haven't already, grab that beautiful coffee, and let's dive straight in.
00:50So when we talk about Ultracoat, one of the best ways to sort of, like, understand or think about this is that we have one boss, so to speak, that's running 10 different specialists. Think of it like a and essentially this boss script is actually running multiple agents in parallel, and then we actually have a judging function.
01:08And this is gonna be one of the biggest growth areas that we're gonna see with AI model use in building anything. Now when you're in ClawCode, you'll notice down on the bottom hand corner, have these different settings. Right?
01:18You have max and you can see it goes all the way down from low, medium, high, extra high, and then max level. Well, that has nothing to do with UltraCode. You actually activate UltraCode by just using the word UltraCode in the prompt.
01:33Think about the reasoning effort as how hard any given agent is thinking. The more that it's thinking, the more tokens that it actually spends. And then Ultracode is more of this new strategy, which kinda takes what we used to do previously to hold on a complete different level.
01:48Again, UltraCode, you can do it for just one task, or if you do UltraCode on, you can run the entire session in Ultra mode. Now, as you know, there's six things that we can do within Cloud Code.
01:57We can have regular chat. We can have plan mode, where we basically plan out the thing we want to build. We can have UltraThink, which extends its thinking over a period of time.
02:06Sub agents where basically Claw just creates mini sub agents to do something. Fast mode which is the same multiple speeder. But an ultra code which is the hero of the show here replaces the managing LLM with code.
02:16And effectively, it's called here a deterministic fan out. What on earth does that action mean, Jack? Well, like sub agents, we're still gonna be spawning up loads of mini clauds, which is the same underlying model, as you can see here, to do things.
02:30But instead of it being an agent that manages it and the performance can decay, this guy in the middle can forget things or whatever it is. We actually have a script running it which means that it will not actually finish until we've achieved the specific output we want to and the general quality of the entire system will be a lot higher.
02:47Though there are so many interesting different approaches that we can use with UltraCode like adversarial verify, a judge panel pipelines, what we call something called loop until dry. I've got a little bit more detail here on what they actually do.
02:59So for example, adversarial verified will spawn a number of independent skeptics, each told to refute a claim and kill it on a majority vote plausible, but wrong findings do not survive the panel.
03:12So this is an example of how you could physically use it. Then I'm gonna show you how you can use this in the Genetic operating system, which is gonna freaking blow your mind. We have this thing called judge panel.
03:20This will generate several independent attempts, have judges score them, then synthesize the winner, grafting the best ideas, and running up. So for example, why don't we show you exactly how that works?
03:30So I could come over here and give it a prompt. I might say, hey there, I'd like you to go ahead and use UltraCode and basically spin up a panel of judges to debate what the best idea is for growing my LinkedIn.
03:43Number one is I learn everything myself. Number two is that I hire a top tier agency. Number three is that I hire an intern to do it and maybe give me a couple different ideas and then just show me what that output is.
03:53So pretty much for random question, but you'll see exactly how it works now. So as you can see now, what's going on here is got a perfect job for a judge panel, which is just one of the applications of Ultracom. It's got a debate, each of your three options gets a sharp advocate, best possible case, then a ruthless skeptic, flaws, hidden costs, and failure mode.
04:11Then we're gonna have a judge, a four person panel is gonna argue this out, each through a different lens, a growth strategist, a CFO, a time leverage coach, and a brand authenticity purist.
04:23Then the verdict, the head judge calls a winner, calls a three and throws in a couple of fresh hybrid ideas beyond your original three. If you're saying, Jack, this sounds like an episode of Law and Order, I've seen courtrooms work like this, you would not be mistaken. So as you can see now, it's got everything here.
04:36Now what's really cool is we have 11 different agents, and crucially, they're not managed by a master agents. We've actually created JavaScript or code to manage this.
04:45Now shut this up. I click on this, you can see, look at what's happening. Our debate is happening before our very eyes.
04:50And we can see we've the judge, no agents have started yet, we've got the verdict. But look at them go guys, I can see the tokens, I can see the tools, I can see how long they're doing it. Now these agents are running in parallel.
05:01We call it parallelization. Right? Try to say that after a couple of cheeky gin and tonics.
05:05Very difficult to do so. The point here is that we basically don't have to wait like an hour. Like things that would normally take sixty minutes, two hours, we could do now so much more quickly because they're all running together in parallel.
05:17Now, it's really cool and actually, you'll often find that agents when you have some questions, not only do they hallucinate, they forget things, but they can become confidently incorrect.
05:27And what that means is you don't know when it's actually telling you something incorrect sometimes. And one of the ways that we can account for this is to have the agents ruthlessly and mercilessly critique each other in a loop. Because the more separate agents that we have challenging one another, the more likely it is to actually arrive at the best idea.
05:47As you can see, now we've got the judges, the growth strategist, the CFO, the time of leverage coach, and the brand authenticity purist. We know from best practices and prompting that the best way to do this is to assign a specific lens, a specific role to actually then attack it.
06:03And this realistically is the best way for you to see exactly how this entire UltraCode system actually works. And by the if this all sounds like I'm speaking Mandarin, I'll put a link down below for the full ClaudeCode master class.
06:15It goes through foundation setup, building websites, power features, memory systems, Hermes agent stuff I have never shared on YouTube. It is the best thing that I've ever built. In terms of a course, you get immediate access to the entire basically, cord code and him, his operating system.
06:31I'll put a link down below so you can check that one out and get all of the juicy benefits because it is very detailed. And you'll understand, guys, when we're using these new strategies, we need to know when to apply it and it's incredible stuff because you need to be slotting in UltraCode, but you've gotta be integrating at the right point.
06:47And I'm gonna show you exactly when it's perfect and when it isn't. And when I show you what you can build with this, we're actually together gonna build out a brand new functionality in the AgenTek operating system that is gonna be incredible, and you'll see what I mean when we get to that point.
07:00One of the cool things I wanna draw your attention to here is we've got the tool use, which is fantastic, how much time it took, but we can see the debate, the judge, and the verdict coming in as it comes through. So it's now currently just synthesizing all of the individual aspects together and it's wonderful.
07:15I mean, all of these are effectively Claude Opus 4.8 and we can run this in any effort level we want to remember. Effort level is one axis and then the ultra code is just simply a different axis.
07:28Now it has essentially completed and now we're gonna get the full output back from Claude. And as you can see, we get the full verdict back. Now this is just one particular example of how we might go ahead and use this ultra code feature.
07:39We can also use this to something called perspective verify, which is essentially basically getting a number of identical reviewers, give each verifier a different lens.
07:47We can do pipeline versus parallel where it lets each item flow stage to stage with no barrier, and a few others including completeness critic, which essentially essentially asks the question, what did we miss? So loads of really interesting ways that we can go ahead and use this.
08:00Now we understand what this incredible technology does, we need to apply it to a project so you can see exactly how this works. So we've got here the ClaudeCode operating system. This is the knowledge graph that I covered in my last video on Graphify.
08:12We're familiar with the operating system giving us the full view of our spend, how much you're spending in each platform codecs, anti gravity, ClaudeCode, dynamic suggestions that dreams overnight based on usage, goals, memory systems, everything. So what we're gonna do here is add in a section to our wonderful Cloud Code operating system or our Autogenic OS.
08:32And what I think we should do, because one of the big areas in the future is gonna be the ability to use the correct model for the correct thing, which is why this OS is so important. Whether you're using Hermes Agent, OpenClaw, or whatever it is that you're physically doing, we need to be using the correct model at the correct time.
08:49And to do that, we need to understand performance and cost. And we call this the best model for the job. In other words, you are model agnostic.
08:57Okay? So we have cost, we have performance, and we have all these different models, Claude, OpenAI, Gemini, Meta. You know, as a company, all by different models, Mistral, DeepSeek, and Grok.
09:06And what's the idea here is that lock in is the enemy of high performance. Okay? The winners are gonna root by task.
09:12So if we're using Homi's agent or even using Claude, if we're doing something with multimedia, we wanna tag in Gemini. Right? If we're doing something like code review, we may wanna bring in ChatGPT 5.5.
09:23But the model landscape can change so quickly and you having to think about it all the time is time that could otherwise be spent actually building things. And so what we're going to do is actually build the system directly in our OS using this UltraCode and I'll show you exactly what I mean. So the first thing I'm gonna do is come over here.
09:38I'm gonna skip it below which is, hey there, I would like you to use UltraCode to create for me an improvement to my ClaudeCode operating system. Specifically, I would like a section, okay, probably somewhere in the home page, maybe at the bottom, that basically covers models that I can toggle on and off.
09:56And effectively, what I wanted to do is have a direct link into various different online sources that essentially shows me what are the best models that are running right now. Maybe we could get some data from OpenRooter. And I would love to see all the different models.
10:10I wanna see all the logos visible. And essentially, I would like it to be able to know what are the up and coming models, what are the pros using right now, what is the performance, what is the benchmark, what is the sentimentality. Maybe it's something we can refresh.
10:23The stated intention here is essentially so that I can look at this and use it as a knowledge base for my Hermes agent, for my Claude code, to know exactly what model I can tag in for any specific task based on price, performance, speed, that kind of thing. Feel free to challenge my thinking on this, but essentially, I want it maybe somewhere on the home page near integrations, automations, that kind of place that doesn't clutter up the interface.
10:50Before we start the ultra think, you feel free to ask me any clarification questions. I do want it to be very up to date. Okay.
10:56So basically, I kind of just explain all the things that I want to. I would recommend that you do this before setting off UltraCode. Because you saw the last example, I just gave it a random idea.
11:06And it brought in a brand specialist, it brought in a, you know, a head of content or whatever it is. Specifically speaking, the Ultra Code system is only as good as the thing it's doing.
11:15It's like, for example, putting a billion dollars into a project about trying to find sand in space. We could spend it. We could put our biggest efforts and war power there, but are we actually putting it in the right place?
11:27You know, answer is probably not for some things, which is why you always wanna have this context conversation ahead of time. So what's cool here is asking us a series of questions.
11:35So what should the on off toggle actually control? In other words, how does this feed Hermes and So essentially, you're gonna come down and give it this feedback here.
11:43I don't need you to send anything to Hermes and ClaudeCode, but effectively, this is going to be a resource that we can call using Hermes or ClaudeCode to make informed decisions. Okay. I'm gonna come down and click on next.
11:54How fresh should the data be given the benchmark sentiment have no clean API? I think we should do something like on a weekly basis. So hybrid live open reader plus dated snapshot, I think that's great.
12:04Footage on the homepage. Cool. So now here's what I think about this.
12:07Awesome. For the homepage, what I think would be great is I wonder if we had see these sessions per day at the bottom? Why don't we just make that something that's toggled, that you can toggle on a couple of things.
12:18Like, maybe we have something like sessions and we can toggle between sessions and models. So we can kind of, like, really think about the information density on the dashboard to make it a pleasure to use so we don't over complicate it.
12:31Little side hack, by the way, that is one of the biggest almost I said fumbles, one of the biggest side steps I see with any kind of dashboard or website in text density. Way too high.
12:41You'll see it on dashboards with like a 100 things on the left. I would never go above five. What I would do as we build this out in the community, I would probably either bring them into one section.
12:50Okay? Agents is slightly different because it's very clipped. The mind clusters things together.
12:55It groups things together. So it's okay looking at five things, but I wouldn't have six. I'd probably merge two together or remove one or do something along those lines.
13:02And as you can see, alt code is now working on the right hand side and we can see the different phases that we've got. We've got recon, so this is gonna do some reconnaissance on it and they can see we've got multiple different agents going on. Then we'll have merge, verify design, and finished by build, which is fantastic.
13:16We've got all these working for us in the background. Beautiful. So now it's gone ahead and designed it.
13:21We can come over to the operating system and take a look at what it has designed, and take a look at this, guys. We can search now by many different things. So we've got down here default, we can search by smartest, what's leaning on the arena, the cheapest, the fastest, the most used based on the benchmarks.
13:38And this is super interesting. So DeepSeek v four Flash is the most used. I can click into that.
13:42I can have a look at different things. It's really freaking cool actually, level of basically configurability we get here. We can filter by frontier models, filter by speed, filter by what's open source and what isn't.
13:53This is really freaking handy actually. I can scroll down and have a look, And at the bottom here, we've got a full breakdown. So as you can see, the DeepSeek v four Flash, this is the number one model in Oprah's and by token throughput, the cheapest credible agentic brand with $1,000,000 contacts.
14:06I've done a full video breaking down DeepSeek, how we use that with Hermes and Claude, and and it's really cool. I just now get this beautiful breakdown for all of the models.
14:14So if I'm ever thinking, dude, what model should I be using for this? I can actually say to Hermes or Claude, it can search the web, obviously, but even if I want context, I'm coming in. I'm like, dude, what are the latest models?
14:25If you like, where's the puck moving? What's new? What's interesting?
14:28And again, guys, if we wanted to build this out and do a lot of the extra details, we can do. Like, that's where we're at with this. We can literally now just get a nice short, sharp, crisp overview of everything that's going on by filtering by the models.
14:40I can search by different ideas. There's mini max. It's just profoundly helpful.
14:44And we bought this using UltraCode. Now there's one thing that we need to understand about UltraCode. But before I do that, I just wanna touch briefly on what we've done here.
14:52So this is an example of an app, an agentic operating system I share in my community, of course. But you'll know, many of you know that I'm building my current text to speech startup, Glider. One question I always get in my community in comment section is around compliance and risk.
15:05I wanna just call this out very quickly because it's really important to understand. And that's it if you're building any apps for yourself, for companies, whether it's an OS or whatever it is, you need to lock down your security and compliance. People always ask me who I'm using.
15:18Company that I use is Vanta. We're using these with Glider because these guys offer over 35 different security compliance frameworks, GDPR, SOC, ISO 27,001.
15:29And I've actually found, honestly, with a lot of, like, enterprise stuff, if you don't have this level of, like, security, they basically don't really it's almost like you don't exist for for many of them unless you've got the level of security and detail. Obviously, everyone needs different things. The reason we use these guys, these guys basically help you automate your security and compliance.
15:48Obviously, there's, like, a million different companies. We use these ones. I actually spoke to them and said, dude, I do a lot on apps.
15:53I should I'd actually like to talk about you on the channel. And they said, cool. They'd be happy to sponsor it.
15:57So I wanna give them a shout out in the video because we find it so important in building beautiful Glider. And this itself does lose very nicely onto what we're talking about the product. Right?
16:06The fact that trust is the product. We've covered some of these different aspects here. And it leads us nicely onto this idea of when we're actually using UltraCut.
16:14Okay. So essentially, we can do it surgically, so we can use it for one turn, or we can have it for an entire session. Now bear in mind, Ultracode is gonna burn tokens like no other.
16:25So you don't wanna be using Ultracode twenty four seven. You're basically gonna burn through your entire credits. You need to be surgical and apply it at the correct moments.
16:34And I'm gonna explain what some of these use cases are in a sec. But on the surgical option, basically, you drop the keyword in and a single request, just say, ultra code, and it will basically go ahead and do it. And after that, you're back to normal.
16:45Or you can do UltraCode on, which is their max mode, a standing default for the session. Every substantial task authors and runs will work for by default. Maximum thoroughness token cost stops being the constraint.
16:56Session only, it resets when you exit. But the tilde r here is the ultra is a power tool. Leave it off for the conversational replies and trivial edits, or you'll pay for orchestration that you quite some didn't need.
17:06Said another way, it's like for So if you're looking for a realistic split when you're doing big builds, you know, probably no more than 20% of the time will I be using UltraCode. Essentially, you know, UltraCode earns its keep on research audits, multi perspective work, and separatable systems anywhere where you don't know the answer's shape.
17:29Good way to think about it, other 80% is back to normal using everything else. People will tell you, and I've seen it online as well, but my wallet is crying watching the video. So it's very powerful, but you don't wanna be using it all the time for things where it's not required.
17:42And a few decision makers to make it easy for you. First of all, you're have four to seven x token vault with this sort of stuff because parallel fan out burns tokens fast. You've got multiple agents working on it.
17:51If I don't need a team to figure out the question, we don't use it. Spin up eats small tasks. It does for anything in the five minutes.
17:57Setup plus verification costs more minute saves, and you can't parallelize a check. In other words, if I have agent m that's gotta pass something to agent b, then pass them into agent c, I could have a billion parallel agents. It doesn't actually make the process any faster because I still need a than b than c.
18:14So really important to bear in mind. Now when do we go and use it? When we're doing research, audits, multi perspective reviews.
18:19Let's say that you're facing a decision in your life or your business, and you would quite like to have a broad variety of perspectives on something perfect for that. Like, hey, where should I live?
18:30Should I buy a springer's spindle or a carcass spindle? You know, life's very important questions. Adversarily verifying any claims.
18:36Building a separatable system you can use and discover what you don't know the shape of something. And we don't wanna use it for single file refactors, normal feature builds, bug fixes, anything with stripped step by step dependencies and quick sub five minute tasks. So use it for those tasks and you'll absolutely crush it.
18:53And so if you look at what I developed here for example, if you're using it strictly and again time was a constraint and cost was something to be aware of, you'd probably use to think about, hey, spin up multiple agents to debate. What would be the most, you know, useful set of functionalities to help a user make decisions about models based on different things?
19:13And they could debate it. But for example, just building it, you'd probably do something more straightforward. So you can see it's got this band of incredible use cases where it is absolutely awesome.
19:22But the big problem here is that building with UltraCode is just one of the key concepts that you need to understand when using Claude. So the next thing that we have to do is learn all of the Claude superpowers to take our business to the next level, which we're gonna do in this video right here.
The Hook
The bait, then the rug-pull.
Claude Code grew a new mode and almost nobody explained the mechanism correctly. UltraCode is not a harder-thinking Claude -- it is a code-orchestrated factory floor where the manager is a script, not an LLM, and the output is guaranteed to meet a defined shape before the session exits.
Named ideas worth stealing.
01:55list
The Six Modes of Claude Code
- Default chat
- Plan mode
- UltraThink
- Sub-agents
- Fast mode
- UltraCode
Six distinct operating modes in Claude Code. UltraCode is the only mode where a code script (not an LLM) manages the orchestration layer, enabling guaranteed output shapes and deterministic completion.
Steal forExplaining to a team why different Claude Code modes suit different task types
02:47list
The UltraCode Pattern Playbook
- Adversarial Verify
- Judge Panel
- Pipeline
- Loop Until Dry
- Perspective Verify
- Completeness Critic
Six named patterns for structuring UltraCode fan-outs, each with a distinct termination condition and use case.
Steal forChoosing the right multi-agent pattern before triggering UltraCode
17:19model
The UltraCode Decision Filter
Use UltraCode for: research, audits, multi-perspective reviews, adversarial verification, separable systems where you do not know the answer shape. Skip it for: single-file refactors, normal feature builds, bug fixes, strict step-by-step dependencies, sub-5-minute tasks. Budget 4-7x token cost; realistic split is 20% of sessions.
Steal forAny decision framework about when to use compute-heavy AI modes
How they asked for the click.
VERBAL ASK
06:10product
“put a link down below for the full Claude Code master class”
Mid-roll at 6:10 after the first live demo completes. Framed as prerequisite context rather than a pitch -- effective because the demo has just demonstrated complexity that makes the structured course feel genuinely useful.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
