The argument in one line.
Augment Code's new task list implementation improved eval scores from 63.2% to 67.5% by keeping the AI constrained to discrete, user-editable tasks rather than allowing it to drift off-track as context fills up.
Read if. Skip if.
READ IF YOU ARE…
- A developer using AI coding assistants on projects larger than single-file tasks who struggles with the model losing focus as context grows.
- Someone evaluating AI coding tools and wants concrete performance data comparing task orchestration approaches before adopting one.
- A programmer currently using text files or external task managers to guide AI coding and wants a native, integrated alternative.
SKIP IF…
- You're still learning to code or working on tutorials and small isolated scripts where task orchestration isn't yet a real constraint.
- You use Augment Code's competitors exclusively and have no interest in switching tools or understanding comparative positioning.
The full version, fast.
Augment Code's newly-released built-in task list is the strongest in-context orchestration implementation currently available, and the proof is a jump from 63.2% to 67.5% on the reviewer's coding evals after the feature shipped. The mechanism splits the AI-coding field into two camps: external orchestrators like Roo Code's mode that dispatch sub-jobs across separate contexts, versus same-context task managers like Claude Code's todos and Augment's new list that keep one chat coherent while constraining scope. What sets Augment apart is direct human editability � you can manually add, edit, delete, or reorder tasks, flip their status, run them all, or hand the list to a fresh chat, no tokens spent regenerating a plan. The takeaway for you: surgical, list-driven execution keeps models on track far better than open-ended prompts, and every other coding tool will copy this pattern soon.
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 – 00:30
01 · The thesis
Names the hardest problem in AI-assisted coding (keeping AI on track at scale) and makes the bold claim about Augment Code's new task list.
00:30 – 02:12
02 · Two schools of thought: split context vs same context
Whiteboards the landscape — Roo Code's orchestrator mode (split context, sub-agents) vs the same-context approach where one chat manages itself.
02:12 – 03:40
03 · Text task lists and Taskmaster
Walks through how he used to solve this — ChatGPT/Claude-generated text task lists, then bolt-on tools like Cloud Taskmaster (screenshot shown). Powerful but takes setup work.
03:40 – 05:00
04 · Claude Code and Augment Code's built-in task lists
Frames Claude Code's todo list and Augment Code's new task list as the in-flow same-context answer. No orchestration needed — the agent manages itself.
05:00 – 06:20
05 · The receipt: 63.2% → 67.5%
Shows his Best-AI-Agents leaderboard. Augment's eval score jumped from 63.2 (May 30) to 67.5 after the task-list update — a real, measurable boost moving it in line with Klein and Roo Code.
06:20 – 08:20
06 · Live demo: manual add, run-all, status control
Screen-share inside VS Code. Adds a task manually, edits status manually, hits run-all-tasks. Argues this is incredible because you can edit individual tasks without spending tokens to make the AI redo a plan.
08:20 – 09:50
07 · What it gets right (and one nitpick)
Praises that Augment knows WHEN to generate a task list — it picks complex queries and skips simple ones. Nitpick: panel is binary open/closed, can't be resized to a partial state.
09:50 – 10:00
08 · Continue-in-new-chat + import/export
Lists the workflow extras — push a task list into a brand new chat (he uses this), import from markdown (untested), export (untested).
10:00 – 11:05
09 · Reining in the AI + closing prediction
Generalizes the lesson — task lists 'rein in the AI,' which is why Claude Code feels controllable. Predicts every other AI coding tool will copy this pattern because it makes too much sense.
Lines worth screenshotting.
- Augment Code's built-in task list improved its benchmark eval score from 63.2% to 67.5% — a 4.3 percentage point jump attributable to a single feature.
- The ability to manually add, edit, remove, and reorder tasks inside Augment Code's task list gives the developer granular control that Claude Code's plan-approval flow does not offer.
- Task list management is the primary mechanism by which AI coding tools stay on track as context fills — without it, the agent drifts from the intended path as the conversation grows.
- The split-context approach (Roo Code orchestrator spinning out separate subagent contexts) and the same-context approach (task list within one session) represent two fundamentally different architectures for managing complex coding work.
- Surgical, small changes reviewed incrementally produce better outcomes than large sweeping AI edits because the developer can verify each change before compounding onto it.
- Claude Code's built-in to-do list has already demonstrated that task-list management significantly reduces off-rails behavior — Augment Code copying the pattern validates the architectural decision.
- The ability to hit 'run all tasks' after reviewing and tuning a task list combines the efficiency of batch execution with the control of human-approved planning.
- A task list that the AI auto-populates based on context — and the developer can then modify — is more efficient than either pure AI autonomy or pure human planning.
- Eval scores on a standardized benchmark provide a quantitative anchor for comparing AI coding tools that would otherwise be evaluated only through subjective feel.
- The prediction that every other AI coding tool will copy Augment Code's task list implementation is backed by the pattern of Claude Code's to-do list already spreading the architectural norm.
- Claude Code going from frequently doing unwanted things in early 2025 to reliably staying on task by mid-2025 demonstrates how much task-list-style constraints improve practical reliability.
- External text-file task lists (the approach GosuCoder used before Augment Code shipped its built-in version) work but add coordination friction that a native in-context list eliminates.
Steal the format: feature review with a receipt.
GosuCoder playbook
Every JoeFlow / Mod Boss feature ship deserves a number — a before/after metric on a real workflow — not a vibes review.
- Open with the universal pain, not the product. 'The hardest thing in AI-assisted coding is keeping it on track' lands before he ever says Augment Code.
- Whiteboard the landscape first. Show where the new feature fits in a map of all existing options. Makes it feel inevitable, not random.
- Always have a receipt. The 63.2 → 67.5 eval delta is what makes this a tool review and not a tool ad. Joe needs a number for every JoeFlow accuracy/speed claim.
- Predict the future at the end. 'Every tool will copy this' creates an evergreen rewatch hook — when the next tool ships the feature, the video gets fresh relevance.
- Keep the demo inside the editor where the audience already lives. No fancy cuts, no Premiere transitions — VS Code screen-share + face-cam PIP is enough.
Terms worth knowing.
- orchestrator mode
- An AI coding agent feature that breaks a large task into smaller sub-tasks and routes each one to specialized sub-agents or modes — keeping the AI on track by dividing work rather than attempting everything in a single context window.
- task list (AI coding)
- A structured list of steps generated by an AI coding assistant before it begins work, used to keep the agent focused and allow the user to review, edit, or remove individual items before or during execution.
- eval score
- A numerical result from running an automated evaluation suite against an AI coding tool — used to compare performance across tools or versions by measuring how well the agent completes a standardized set of coding tasks.
- Claude Taskmaster
- A popular third-party task management framework for Claude Code that generates and manages structured to-do lists from a project specification file, helping the agent work through large tasks without losing focus.
- PRD (Product Requirements Document)
- A document that defines the goals, features, and scope of a software project — used in AI-assisted development to give the agent a structured brief before generating a task list or beginning implementation.
- context window (coding)
- The total amount of conversation history, code, and instructions an AI coding tool can hold in memory during a session — as it fills up, the model becomes more likely to make errors or go off-track on complex tasks.
- AugmentCode
- An AI coding assistant known for its strong codebase indexing and retrieval capabilities — reviewed here as one of the top-tier tools after shipping a built-in task list feature that measurably improved its eval scores.
Things they pointed at.
02:55toolCloud Taskmaster
03:40toolClaude Code
03:55toolAugment Code
00:30toolRoo Code (orchestrator mode)
05:27toolBest AI Agents leaderboard (his own site)
09:35toolOpenAI o3 (for PRD generation)
Lines you could clip.
00:00
“One of the hardest things to do with AI assisted coding is keeping the AI on track in large coding projects.”
05:20
“Augment Code went from 63.2% in my evals to 67.5 — pretty huge boost very consistently because of the task list management.”
10:00
“It really does rein in the AI. I felt this in Claude Code, and that's probably one of the reasons why I like Claude Code so much.”
10:38
“I would actually be surprised if we did not see things like this in some of the other AI coding tools because it just makes way too much sense.”
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
00:00One of the hardest things to do with AI assisted coding is keeping the AI on track in large coding project.
00:10And I will say probably something kinda bold, but I think AugmentCode might have just released one of the best task orchestrators, task list implementations that I've seen so far in any of the AI coding tools.
00:25And to back up a little bit, we can kind of see this sort of that people are going that these companies are going.
00:33On one hand, we've got the orchestrator mode in RUCODE. What that does is it is it's kind of like a task list if you think about it.
00:43You give it a large thing. You give it a large amount of of work to do. It actually breaks it up and then orchestrates out the smaller jobs to, like, other modes.
00:53For example, you might need to do an architecture mode or you might need to go do some coding. But that actually is a really interesting thing because you're no longer, like, confined to a single context, and you're able to really keep the AI on track by that orchestrator mode, kinda dishing out the work.
01:12Where on the other side, there's we're gonna keep it all in the same context. One model chat log and so on.
01:19And one of the things that I've actually done probably for the last, I don't know, year now at this point, is I found that just small surgical changes, basically, not letting the AI do these big sweeping things, actually, has honestly helped me a lot because I can review the code easier.
01:38I can go in, and I can make sure that the AI doesn't get off track to really being surgical about it. But some other things have evolved over time. So one thing I did for a while was I was using kind of text based task list.
01:53So I would do I would have, you know, maybe ChatGPT or Claw generate a list of things that needed to be done, And then I would use the AI assisted coding tool to actually go through, read that file, and use that sort of thing as what it needed to do.
02:10But then we've had things kind of add on like Taskmaster. So Cloud Taskmaster might be one of the more popular ones. So here's an image of kind of what that looks like.
02:21This thing's actually pretty sweet, honestly. But at the same time, I would say, like, it does take a little bit extra work to get it set up.
02:28So we've got these really cool, really great implementations that basically allow you to orchestrate and keep the AI on track.
02:37Because if if any of you know, like, if we can constrain the AI to do our bidding, we can actually accomplish some pretty amazing things with it.
02:47And a lot of people talk about the big complications with AI, and I think one of the biggest complications with AI is just making it stay on track. And as the context fills up, its likelihood of staying on the path that you want it to be on kind of a little bit.
03:05And AI has gotten a lot better at this from where it was even, let's say, eight months ago to today. It's gotten a lot better. So that that's where Cloud Code and Augment Code have kind of come in.
03:17And there may be others too, but these are the two I know the most. Cloud Cloud Code has this built in to do list function where you can actually get it to generate its own task list so you don't have to orchestrate anything.
03:31It just manages it, and it keeps itself on track. And what AugmentCode has done is it has also just released its own version of the task list.
03:43And this thing is freaking awesome. And I'm gonna go through some of the features on it, but the big thing I wanna cover is it went from 63.2% in my evals to 67.5.
03:57And I've actually got an early version of my site up if you go to the best AI agents. You can see here if we search by CloudForms on it, the last time that I had ran AugmentCode was sixty three point two o, and that was on May 30.
04:12The task manager alone has improved its scoring quite substantially right in line with Klein and RootCode, know, maybe slightly behind that.
04:21Still number seven spot, but it is all relatively within margin of error when you see, like, a fraction of a point difference between them. So pretty huge boost very consistently from claw or from AugmentCode because of the task list management.
04:39Now they just make a huge difference. Ever since ClogCode, and I've been working with that, I very rarely have Cloud Code go off the rails.
04:51Where if I go back a few months ago, some of you might have known that I was venting a little bit about Cloud three dot seven because it would just go do things I didn't ask it to do. I don't have that problem anymore.
05:03And in fact, I would say clog ClogCode, in particular, has kept my my, um, what I'm actually executing, like, really, really well constrained. And I've just totally enjoyed working with it because I don't have to worry about it doing things I don't want it to do. But let's talk about AugmentCode a little bit because what what I said before is I think they have created one of the best, if not the best implementations for task list.
05:29So one thing to note is not everything is going to actually generate a task list. So in this particular one, I was actually had it debugging an error message for me. It did not generate a task list here.
05:41But if I go into some of my past ones let's take a look at this one. I think this one may have.
05:47It did not. Okay. So let me go back one more.
05:49K. Here's one that actually did. So this one actually generated a task list of things that it wanted to do.
05:56And I'm just gonna stay here for a second and kinda show some of the functionality that I've really enjoyed. So the first thing is you can just manually add a new task. And this is incredible, in my opinion, because I can come in and be like, I want this to do x y z.
06:13And I've used this quite a lot over the last day and a half. Because sometimes I don't even I like what the AI did or I wanna actually come in and I wanna change this to something else.
06:25I can do that without having to actually have the AI generate and use tokens to make that happen. I can go in and actually edit it myself. Now once the task list is in place, you literally can hit run all task, and it will just run through them.
06:39The other thing you can actually control is the status yourself, which I thought was really interesting. It will auto get updated as well as it actually gets completed, but you can also come in and you can say, oh, I don't wanna do that one.
06:54That one's complete, or you can just go ahead and delete them. So really, really very, very cool because when you think about clogged code, for example, you know, you get, like, this plan, for example, that I could go ahead and approve or not.
07:08I can't really change individual things very easily.
07:14So for example, if I if I like phase one, but I wanna change something in phase two, the way I need to do that is communicate back to it to have it make that plan. And that has been fine for me.
07:25Like, honestly, I I haven't had a lot of issues or concerns with doing that. But as I've started using all of my codes new task list, I've realized how much nicer it is to be able to come in and actually tune the task itself, remove a task, add a task.
07:43So this thing is incredibly powerful. And what I would say is, um, if you want it to generate a task list, it seems to do, I would say, a really good job of finding the right times rather to actually use a task list, and it finds the right times when it shouldn't use a task list.
08:02Because I've I've had somewhere when I put in a query, it knows that it's complex enough and it's like, boom.
08:09Here's my eight things. I'm gonna go ahead and do it. And it did that with every one of my evals that I put in.
08:15Where before, I think it scored poorly because a lot of times the agent would just kinda give up.
08:22Give up is the wrong word, but it wouldn't complete it to the fullest extent that it needed to. And, the task list is to accomplish the particular goal that we actually have here. I would highly recommend giving this a try.
08:36It's this little icon here right beside the file change. Now the one thing that I would say is I I sorta wish I had a little bit more control on the size of this thing.
08:47I it's just either open or close, and I would love to be able to, like, bring it down to, like, two or three show. I know this is a minor nitpick, but I do keep trying to, like, drag it up and down this.
09:00And I want this a certain size, and I want this a certain size, kind of a minor a minor issue. But it is just worth kinda calling out.
09:08The other thing that I would say is you can actually take your task list and put it into a new chat. So I've done this a little bit where I was actually planning through something. And then then now that I have my task list, I just started a brand new context window with it in a brand new chat.
09:25I have not tested the import from markdown yet. I am very interested in checking that out because you could go in, you know, o three, for example, have it help you generate the PRD or the technical task list that you want and bring it in, and that could be kinda sweet as well.
09:43And I haven't really played around with exporting yet because I haven't really had much of a need to. But continuing new chat is the one that I've I've used a couple times at this point now. So, anyway, I just wanted to get kinda touch on this real quick and just share that massive improvements on evals with this.
10:01It really does rain in the AI. I felt this in Clog code, and that's probably one of the reasons why I like Clog code so much. I hadn't quite put that, you know, that into into words yet because Cloud Code is so controllable.
10:15I've said that. It's so controllable. And I think a lot of it has to do with the way it kind of breaks out its own tasks that it's doing.
10:23And I think Augment code, now that it has this ability, it's just gonna become a lot more powerful for people. Combined with its amazing context engine, , and I'm excited to kinda see what they go next because I was not expecting something like this.
10:38And in fact, I would actually be surprised if we did not see things like this in some of the other AI coding tools because it just makes way too much sense for the way they have this implemented. Really great implementation.
10:50Really excited to kinda see what they do next. Anyway, I'm gonna wrap it up there. Let me know what your thoughts are below.
10:55Have you had a chance to try this out? If not, you should just definitely go check it out because this thing is freaking awesome. Till next time, everyone.
11:02Have a wonderful day. Peace out.
The Hook
The bait, then the rug-pull.
The opening is a thesis statement masquerading as a casual review. Inside the first 18 seconds GosuCoder names the central problem of AI-assisted coding (keeping the agent on track in big projects), then makes the bold claim that Augment Code's new task list may be the best implementation he's seen — which is the entire rest of the video unpacked.
Named ideas worth stealing.
00:30model
Split Context vs Same Context
- Split Context — orchestrator mode dispatches sub-jobs to specialized modes (Roo Code, Claude Subagents)
- Same Context — one chat manages its own task list in-flow (Claude Code todo list, Augment Code task list)
The two architectural approaches AI coding tools are converging on for keeping agents on track in large codebases.
Steal forAny 'state of the industry' creator video — pick the axis, place the players on it, name the trade-offs.
01:40list
Five Things That Help Keep AI on Track (whiteboard list)
- Small surgical changes — don't let the AI do big sweeping things
- Use text-based task lists and have the AI work through them
- Add-ons like Cloud Taskmaster
- Claude Code built-in todo lists
- Augment Code built-in task lists
The whiteboard slide that anchors the whole video — the progression of techniques he's used over the last year.
Steal forAnchor a feature-review video with a 5-bullet whiteboard slide that shows the progression of solutions to a single problem. Frames the new thing as the natural next step.
How they asked for the click.
VERBAL ASK
10:45next-video
“Have you had a chance to try this out? If not, you should just definitely go check it out because this thing is freaking awesome. Let me know what your thoughts are below.”
Soft CTA — three asks bundled (comment, try the product, implied subscribe). No hard sell, no affiliate pitch despite a Scrimba affiliate link sitting in the description. The product itself is the call to action.
FROM THE DESCRIPTION
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
