Anthropic Says Delete Your Subagents

Author: Rob Shocks - Switch Dimension · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Anthropic Says Delete Your Subagents · Channel: Rob Shocks - Switch Dimension · 09:43 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Claude Code

Summary

Anthropic’s guidance suggests that, as models improve, teams should reconsider complex agent harnesses and fold unnecessary subagents into a single agent with skills. The video explains how to test this change by establishing a baseline, converting a subagent’s instructions into a skill, and keeping the subagent only if the folded version performs worse. Subagents can still be valuable for context isolation, restricted permissions, parallel research, and tasks that benefit from separate contexts. The broader recommendation is to evaluate harness components rather than assume that either adding or removing complexity will always help.

Key points

Anthropic reports that one agent with skills often outperformed both a giant prompt and a team of subagents, while also being faster and cheaper in some cases.
Agent teams can mirror company org charts, creating extra handoffs, lost detail, and increasingly complex system prompts—the speaker refers to this as a “Frankenprompt.”
Progressive disclosure lets an agent load a skill’s full instructions only when needed, helping keep the context window focused.
As models improve, scaffolding built for older models may stop helping or may distract the model; complexity should be retained only when it demonstrably improves outcomes.
To evaluate folding a subagent into a skill, test varied prompts, record performance, convert the subagent prompt into a skill, and rerun the prompts in a fresh session.
Keep subagents when they provide useful context isolation, restricted tools or permissions, parallel research, or help with noisy side tasks.
Subagents may be a poor fit for tightly dependent steps, rapid back-and-forth, or building components that must coordinate closely; they can also consume many more tokens.
Skills and instruction files can impose unnecessary work, so review whether the current model would follow each instruction without being told and remove obsolete or unused items.
Shrinking a harness does not always reduce costs or improve results; test changes on your own setup, since the guidance is based on observed patterns rather than a universal guarantee.
The suggested direction is to put work requiring the same memory in one main agent and use workers selectively for parallel or isolated tasks.

Chapters

Start	Chapter	Summary
00:00	Why Anthropic is questioning subagents	The speaker introduces Anthropic’s recommendation to replace unnecessary subagents with one agent and skills, and explains the trend toward simpler harnesses.
01:00	How agent harnesses become complicated	The video describes org-chart-like agent teams, the cost of handoffs, progressive disclosure, and why stronger models may need less scaffolding.
04:04	How to fold and test a subagent	The speaker outlines a baseline evaluation, converting a subagent prompt into a skill, and comparing results before deciding whether to keep it.
05:07	When subagents and skills still help	The video covers useful cases for subagents, including context isolation and parallel research, while warning about token costs and coordination problems.
06:38	Cull the harness, but evaluate the trade-offs	The speaker recommends reviewing skills and instruction files, notes that some domains still need guardrails, and closes with a checklist for testing harness changes.

Quotes

one agent with a set of skills was frequently beating both one giant prompt and a team of subagents on quality. — 01:00

you should only add complexity when it really demonstrates it improves the outcomes. — 02:32

keep the sub-agent only if the folded version scores worse. — 04:34

The biggest win for sub agents is parallel research. — 05:37

the harness is shrinking as the model grows. — 09:12

Concepts and entities

Anthropic, Claude Code, subagent, skill, Harness, Frankenprompt, progressive disclosure, Conway’s Law, Claude Opus, Cursor, Grokbot, Cognition, Microsoft Research, Gedai Cats

Full transcript (original language)

00:00 Subagents are great. Subagents are bad. It is kind of hard to keep up. Anthropic just told companies to get rid of their unnecessary subagents and fold them into one agent with skills. And this kind of reduction is a trend we’ve been seeing over the last few weeks and months. We’ve seen people reducing their agents.md files and also calling their skills. The fact here is the models are getting better and the complicated harnesses we’ve been developing around them are not required in every instance, at least. You can actually see this with

00:30 Claude code itself. They have cut the system prompt by over 80% for their stronger models. Notice the way I said stronger models, and that’s because this isn’t going to work for every type of model strength. So, with this new research and guidance from Anthropic and others, should we be turning off our subagents as well? I’m Rob from Switch Dimension. I’m going to go through what Anthropic said, why all this is happening now, and I’m going to show you how to apply this yourself when it comes to your skills, your Claude.md, and your subagents. So, this is all coming from Anthropic’s Applied AI team, and it came

01:00 out on the 22nd of October. So, they’re saying one agent with a set of skills was frequently beating both one giant prompt and a team of subagents on quality. It was often cheaper and faster, too. And who doesn’t want fast and cheap? In one case, a customer tested their multi-agent setup against a single loop, and the single loop ran about twice as fast. So, how are we ending up with so many subagents in the first place? Well, when it comes to software factories and running processes for companies, each team is building out its own set of agents. And then someone

01:31 folds an orchestrator on top to manage all these agents. And before long, you’ve got this agent team that is a copy of your org chart. We saw this a lot with open cloud, people were building in hundreds of agents, and it just got way too complicated. So, this is called Conway’s Law. So, the problem here is when agents are communicating, every handoff loses detail, and then every team or every agent you’re adding is going to add a system and a little bit more complexity to the system prompt. Anthropic actually have a word for this called the Frankenprompt. So, the fix for this is progressive disclosure and we’ve seen this baked

02:02 into how skills have been built for the last couple of months. So, the agency is a short description of the skill and tools and then only loads the full thing only when it needs it. And this is a big part of how they cut down Claude code’s prompt. So, essentially why would you load in all of the information if it’s not necessarily going to be needed? Just give an indication that it’s there and the agent can go look it up when it’s required and that keeps the context window nice and tidy. Now, Anthropic don’t say sub agents are dead and I am going to come back to you on why you’re still going to want one. So, why is all

02:32 this simplification happening now? A lot of what we build around the model like the sub agents and these long system prompts is scaffolding for the models that we had like 6 months ago. With a lot of the newer models we’re getting all that complexity we’ve added has kind of stopped paying off and in fact it can actually be quite dangerous and distract the model from what it needs to be doing. Anthropic are saying you should only add complexity when it really demonstrates it improves the outcomes. So, Gedai Cats who worked on Cursor’s agent harness team and now on Grokbot calls this the flywheel. He’s a great

03:03 talk on this on the AIE channel you should check it out. I’ll leave a link in the description. So, his point is we’ve built these harness then the models actually learn from the harnesses we’ve built and it makes harder tasks possible and when the models catch up with the harnesses we can actually delete what’s gone before and grow a new harness around more advanced models. It just goes to show we’re at a stage where our systems are never going to be at rest. They need to be continually updated. He points to two studies where the harness barely changed coding scores on newer models but the cost per task varied a lot. One cool tip I got from

03:34 that talk, his team logs every temporary guardrail each time it fires. So, basically they’ve put these little hooks and guardrails in place to keep the model on track. They call them shims. And when these shims start firing, they actually have another bot which opens up a pull request to delete that shim. In In case, a sub agent is basically one of these shims, and once the main agent can do its job, you can just delete it. Now, having an agent like that do the cleanup of your harness is an advanced step, but you can see where we’re going with this. This idea of a self-evolving harness.

04:04 So, how do you fold a sub-agent? Well, Anthropic says start with the sub-agents that chain on each other’s output, the ones that are touching the same data or get called together. Now, if a sub-agent exists because its tools are locked down to a certain process, you want to leave that alone. Restricted tools are a reason to keep it. So, first of all, get your baseline. Use 10 varied prompts that should fire the sub-agent and run each one twice. Track whether the right sub-agent actually gets called with your prompt and whether it actually passes the test. So, when it actually comes to folding it in, in

04:34 Claude code, a sub-agent file is a description plus a system prompt. Remember, this is all just natural language. And then, a skill is its description plus some instructions. So, basically, the sub-agent prompt becomes the body of a new skill.md, and then you rewrite the description to say what it does and when to use it. Essentially, you’re getting rid of the sub-agent and just folding its functionality into a skill. So then, move the old sub-agent file aside instead of deleting it and then rerun the same prompts in a fresh session. And then, keep the sub-agent only if the folded version scores worse.

05:07 Now, all of this sounds really time-consuming to run and to test, but you can set up an agent to go and actually run these tests for you. I would actually just feed it the Anthropic article or even this video and ask it to run the analysis on your sub-agents. So, when does a sub-agent still earn its place? So, Anthropic gives us three cases. So, the first scenario is with context bloat. If you don’t want your context window to get overloaded by all the research and content that’s being flooded in, it might measurably hurt the main agent. What you can do there instead, of course, is use parallel sub-agents

05:37 who’ve got their own fresh context window, and you can throw them away, and they will summarize their findings and give it to the main agent. Another case is where one agent keeps picking the wrong tools. So, in that case you might split agents by context, but not necessarily by org chart. The biggest win for sub agents is parallel research. In this case, Anthropic’s multi-agent research system beats single-agent Claude Opus by 90% on their internal evals. Sub agents are also great for searching code bases and noisy side tasks like digging through logs and reviewers that are only going to read

06:08 content. Of course, the downsides here is you’re going to use a hell of a lot more tokens in this process, and when it comes to writing or building something, use of sub agent isn’t always the best case. In an example from Cognition, they had one sub agent build a Flappy Bird background that looks a lot like Super Mario, and then another sub agent went and built a bird that did not match the background whatsoever. So, if you have steps that depend on each other, there’s quick changes, lots of back and forth with you, then keep them in the main agent. More on culling your harness.

06:38 You’ve probably already heard about deleting your skills. That advice has been around for a few months, and it kind of follows the same logic. Every skill description and every line in your Claude that MD and your agents.md is harness. And a lot of it might have been written for an older model. And you’re going to be paying for all this. Claude codes docs say every skill in the listing adds to your context on every single turn, whether Claude codes is using this or not. As a starting point in one of my courses coming up, I recommend that you absolutely cull your harness and start from scratch. So,

07:09 skills can also make busy work. A paper from Microsoft research and two universities found that what ends up happening with skills is you’re turning these validation checklists and construction recipes into mandatory work. If your skill says always run the full test suite, it might do that even if there’s like a one-line change to a button in your code. You need to give the model a little bit of flexibility to decide when that’s going to be overkill. A handy skill around this is {slash} skill doctor. You can run it yourself in Claude code, and And to delete the

07:40 skills that never fire. And then go and read your Claude.md file or your agents.md file and ask yourself, would the current model I’m using do this anyway without being told? And if yes, just cut it out. And if you’ve time, of course, just run the eval after to test. Now, it’s not straightforward. On the other side, harnesses don’t always need to shrink. For an example, with Grok Bot, it will run a coordinator bot and then have a fleet of workers. And because these bots are doing things that fall outside the realm of coding, where

08:11 our models have gotten really good, they’re actually adding certain guardrails back because these are areas that the models just really haven’t nailed yet. Now, as a counterpoint, also trimming skills won’t always fix your bill and reduce it, either. A July paper shortened skills by up to 54% and the cost per task didn’t necessarily drop. In some cases, skills are small, mostly cached parts of what the agent reads. Cached meaning that when they’re brought in in the initial prompt, you’re charged a lot less for them in subsequent turns of the agent. Also, Anthropic’s guide

08:41 here is just a pattern they’re seeing from their customer work. There’s no pass rates or sample sizes, so we can test it on our own setup. We just have to feel our way through this ourselves. So, to give you a solid checklist in your coding agents or your system, what I would do is I would build a small eval or just just get your agent to build it and then fold in one sub agent at a time. By that, I mean basically you’re turning the sub agent off and just taking what it did and putting it into a skill. And then keep that sub agent only if the folded version scores worse. All going right, what we should be seeing is

09:12 the harness is shrinking as the model grows. Where sub agents are still going to be useful is for isolation, permissions, different things like that. And then, of course, for parallel reading and research. And let’s try and get to a place where we put everything that needs the same memory in one agent. We’re kind of seeing this with the likes of Grok Bot and Juwel. We’ve got this main agent that spins off lots of different workers. And inevitably, I want to be in a situation like her where I’m just chatting to one agent to get all of my things done. I don’t to multiple different agents that I have to go to to decide, well, what are you for?

09:43 What do you do? I just want to be able to work with one agent, my chief of staff. If you want to go deeper on building agents like this, check out the switch dimension course and community in the description. I’d love to know how everybody else uses sub-agents. Have you got a ton of them doing various different tasks or do you prefer to just work with one agent and a lot of skills? If this was useful, subscribe. If you want to learn more about Anthropic and their software development life cycle, you can check out this video here.