00:00:00Are we going above managing agents and giving that role to AI too?
00:00:03Ever since AI entered the agentic space and started interacting with tools,
00:00:07everything has changed. Now we let AI interact with tools on our behalf,
00:00:11using agentic systems like Claude Code to do the work for us. Our role has shifted to simply
00:00:15delegating tasks to agents and letting them handle the execution. But we are already moving beyond
00:00:20this delegation. Claude has a new update where it handles the tasks in a different way than it
00:00:25usually did by taking over much of the delegation itself and integrating it directly into the
00:00:30product. This added another layer of abstraction and changed how we work. This is exactly what this
00:00:35startup founder talks about in his article. Now Claude's new update might not sound like
00:00:40something impactful because apparently it sounds like new to-dos, but it's actually a huge update.
00:00:44The main idea behind the agent swarm is having multiple AI agents coordinate on complex tasks,
00:00:50spawning sub-agents and managing dependencies in parallel. This means they can take a complex task
00:00:55from a user and break it down into multiple tasks for AI agents, letting them work in isolation.
00:01:00So now you can talk to Claude as if you are talking to a project manager, giving it a wide task to work
00:01:05on, and Claude automatically does the breakdown and delegation. With this update, your task can
00:01:10survive the clear command and even a session restart. We'll explain exactly how that works
00:01:14in just a moment. Before this task system, when working with Claude, we had to hit compact more
00:01:19often because even if it did divide tasks, in the end it was still a single brain trying to hold
00:01:24complex processes in its small, limited context window. This became more annoying when working
00:01:30on larger tasks because it used to lose context more often, and we had to create workflows with
00:01:34structured notes so it wouldn't lose context as frequently. Now we've noticed that when working
00:01:39with Claude, we don't need to hit compact as often as we used to. What we used to do manually with
00:01:44notes in Claude.md or other guidance files, they've now incorporated into their own product. The agents
00:01:50are not sharing a single context window. Each agent actually has its own context window. As we mentioned
00:01:55earlier, you interact with the main Claude, who acts as a task coordinator. This coordinator creates a
00:02:00task graph that identifies and breaks the work down into smaller tasks. It then determines the type of
00:02:06each task, whether it's sequential, meaning the previous task needs to be completed before starting
00:02:10the next one, or non-sequential or parallel, meaning there are no dependencies and they can run at the
00:02:15same time. Each task follows a full workflow to investigate, plan and implement the task, with each
00:02:20stage being blocked by the previous one. Once the task graph is created, it spawns agents and delegates
00:02:26different models to each task based on its complexity. Some tasks, like exploring folders, don't need heavy
00:02:32reasoning from Opus 4.5 and can be handled by Haiku or Sonnet models. Each agent gets a fresh 200k
00:02:38context window, which is isolated from the other processes. This is different from how Claude worked
00:02:43before, where it relied on a single context window, which caused problems. With this system, each agent
00:02:48is able to focus on one thing. You've probably noticed we build a lot in these videos. All the
00:02:53prompts, the code, the templates, you know, the stuff you'd normally have to pause and copy from the screen,
00:02:58it's all in our community. This video, and every video before it too. Links in the description.
00:03:02Now that was the detailed explanation of how the new task system works, and at first it might not sound
00:03:08much different. Previously, it used to write tasks into the context window, and once the context window
00:03:13filled up, it had to compact, which caused the to-do's to get messed up in the process. Now, tasks aren't
00:03:18just in the context window. They've added a new task folder inside the main.claud folder, where there's
00:03:23a folder for each session, identified by the session ID of that session. Inside each folder, there's a set
00:03:29of JSON documents representing tasks in the system. These JSON files are identified by their IDs and
00:03:34contain a name, description, and status. The two main keys to focus on are blocks and blocked by. The blocks
00:03:41key lists the tasks that are blocked by the current task, while blocked by contains all the tasks that
00:03:46are blocking the current task, and after those execute, the current task can proceed. This setup
00:03:51ensures the correct sequence because it creates a dependency graph showing which tasks depend on
00:03:56others and which are blocked. Basically, this guides claud so it can't skip a task until the required one
00:04:01is completed. Without this graph feature, you would have had to explain to claud again every time you
00:04:06wanted to use the clear command, but that's no longer necessary. This logic has been externalized
00:04:11into a file structure, which allows the system to retain its state even when the session ends, no
00:04:16matter how many times later you come back to it. That way, claud doesn't have to figure out which
00:04:20tasks to redo. The graph doesn't forget and doesn't drift from what it needs to do. The folder names are
00:04:26currently just random IDs for the session, but if you set an environment variable with a custom name,
00:04:31it will identify the session by that name. This ensures that tasks aren't lost even if you close
00:04:36your terminal, and claud can continue the session seamlessly. With this update, Anthropic has finally
00:04:41killed the Ralph loop, which was originally all about re-anchoring the task system. Now, claud
00:04:45handles it automatically on its own. Also, if you are enjoying our content, consider pressing the
00:04:50hype button because it helps us create more content like this and reach out to more people. Now, this
00:04:55approach matters because it gives claud a degree of freedom in parallelism by effectively managing the
00:05:01parallel and sequential steps together. Claud identifies everything that can run in parallel
00:05:06and everything that cannot, and based on that, it saves time in completing the tasks. For example,
00:05:11it sees that task 1 and task 2 have no dependencies, so it spawns both at once. At the next layer,
00:05:16it identifies that task 3 and task 4 are blocked by task 1, so it waits for task 1 to complete before
00:05:22starting the next tasks. In this way, the last task completes in just three cycles. Previously,
00:05:27these 5 steps would have taken 5 waves, each waiting sequentially for the previous one. But
00:05:32with this approach, execution time is reduced by running tasks simultaneously. This not only saves
00:05:38time, but also reduces costs, because the model matches its effort to the tasks and doesn't waste
00:05:42extra tokens on smaller tasks. But before we see them in action, here's a word from our sponsor,
00:05:47Lovart. Looking at these designs, you'd think a pro agency made them, but this is the first
00:05:52AI design agent built with true creative intuition. Design is easier with Lovart because it helps you
00:05:57visualize any concept instantly. From complex packaging and interior layouts to unique jewelry
00:06:02collections, it's the design agent that delivers professional creative work to get the job done.
00:06:07The real power lies in its exclusive editing features. Usually, AI text is a mess,
00:06:12but with TextEdit, I can rewrite headlines perfectly just by typing. With Lovart AI,
00:06:17you can generate stunning posters for work and use edit elements to move, adjust, or swap individual
00:06:22layers or touch edit to swap or change objects precisely without breaking the style. This lets
00:06:27you produce way more high-quality posts without extra effort. You can even turn the final static
00:06:32visual into a video with one click. Start designing for free by checking the link in the pinned comment.
00:06:38Our team tested this swarm across multiple scenarios on both Claude Code and Co-Work.
00:06:42For those who don't know, Co-Work is basically Claude Code, but for non-developers. The idea
00:06:47comes from the fact that when they first developed Claude Code, it was intended for developers only.
00:06:52But they realized it could be useful for almost everything else. Co-Work has more guardrails than
00:06:57Claude Code because it's not aimed at developers. This helps prevent the agent from accidentally
00:07:02deleting or messing with something it shouldn't, making it much friendlier for non-technical users.
00:07:07Our team has also been using it for non-development tasks like research, planning, and even managing our
00:07:13channel's ideation process by connecting it with Notion. So Anthropic made it simpler and released
00:07:18Co-Work, which essentially does everything Claude Code does, interacting with file systems and making
00:07:23changes when needed. Co-Work works really well if you want to organize folders or make changes
00:07:28in them. We've been using Co-Work extensively for this purpose. We had a folder with a lot of
00:07:32projects, mostly for testing purposes, and we were having trouble navigating it to find a particular
00:07:37skill we had used in a previous project. So we asked it to create a document detailing what each
00:07:42project contains. We also asked it to look at the Claude.md and the reusable commands we'd created
00:07:47and differentiate based on that. It started by exploring the folder we had connected and creating
00:07:52to-dos. Then, it used the same agent swarm method we talked about earlier with Claude Code. It spawned
00:07:58multiple agents to read the files in batches and create documentation for what each project contained.
00:08:03In the end, every project had a file summarizing what it does, making it much easier to navigate
00:08:08and find exactly what we needed. We used Co-Work for feasibility and market research for an app
00:08:13we were working on, and it created a proper document containing all the findings. Just like
00:08:18Claude Code, it asked questions and, based on the answers, produced a comprehensive report. It saved
00:08:23the report in the folder we had connected Co-Work with. You could do something similar with Claude
00:08:27Chat, but now it actually has access to the documents inside the folder, which helps guide
00:08:32the research much more effectively. The generated report also had proper formatting because Co-Work
00:08:37comes with specialized skills to create documents better than before. Now, once the research and PRD
00:08:42documentation was complete with Co-Work, we moved to Claude Code for the actual implementation part.
00:08:48We asked Claude Code to look at the document inside the folder, which was used to guide Co-Work on the
00:08:53project idea for which it did the research, and break it down into different components, focusing
00:08:57on one aspect of the PRD. It analyzed that the PRD contained multiple sections and realized that these
00:09:03could be handled in parallel since they were not dependent on each other. So, it spawned multiple
00:09:08agents to work on writing them simultaneously, with each agent working independently. Without
00:09:13the parallelism, it would be 16 sequential steps which were reduced to one step because of
00:09:18parallelism leading to significantly speeding up of the process. Now, Claude breaks down complex tasks
00:09:23automatically, but sometimes it doesn't because it does not consider the request to be complex enough
00:09:28for breakdown. If it doesn't, you can prompt it with something like "break this down into tasks with
00:09:34dependencies". It will then create the dependency graph and use it to manage the workflow. You can
00:09:38even see the todos by hitting Ctrl+T. Since this was a long-term project, we set the CLI flag to
00:09:44the project's name so we could return to it later. That brings us to the end of this video. If you'd
00:09:49like to support the channel and help us keep making videos like this, you can do so by joining
00:09:53AI Labs Pro. As always, thank you for watching, and I'll see you in the next one.