0:00Sub-agents are great; sub-agents are bad — and Anthropic just told companies to get rid of the unnecessary ones and fold them into one agent with skills. It's part of a wider trend: people reducing their agents.md files and culling skills. The reason: models are getting better, and the complicated harnesses we built around them are no longer required — at least not in every instance.
The guidance
0:57From Anthropic's Applied AI team (Oct 22): one agent with a set of skills frequently beats both a giant single prompt and a team of sub-agents on quality — and is often cheaper and faster too. In one case, a customer's multi-agent setup was ~2× slower than a single loop. The supporting evidence: Claude Code has cut its system prompt by over 80% for its stronger models — "stronger" being the key word, because this doesn't work for every model strength. Note the framing: Anthropic isn't saying sub-agents are dead — just that they should earn their place.
Why sub-agents proliferate
1:25In software factories, each team builds its own agents, then an orchestrator sits on top, and before long your agent team is "a copy of your org chart" — Conway's law. It got out of hand with OpenClaw, where people built hundreds of agents. Every agent handoff loses detail, and every team adds a system prompt plus complexity — what Anthropic calls the "Franken prompt." The fix is progressive disclosure: the agent sees a short description of a skill/tool and loads the full thing only when needed — "why load all the information if it's not necessarily going to be needed?" That's how Claude Code slimmed its prompt.
The flywheel
2:32Why now: most of what we built — sub-agents, long system prompts — is scaffolding for the models we had six months ago. With newer models, that complexity has stopped paying off and can actively distract. The mental model (from Jediah Katz, ex-Cursor agent harness, now on Grokbot): the flywheel — we build harnesses, models learn from them, harder tasks become possible, models catch up, and we delete the old harness and grow a new one around more advanced models. "Our systems are never at rest." His team logs every temporary guardrail — they call them shims — each time it fires, and when a shim stops firing, a bot opens a PR to delete it. A sub-agent is basically one of these shims: once the main agent can do the job, delete it. That's the "self-evolving harness."
How to fold one
4:05Anthropic's starting point: fold the sub-agents that chain on each other's output — the ones touching the same data or called together. (Leave alone any sub-agent that exists because its tools are locked down to a process — restricted tools are a reason to keep it.) The method: (1) get a baseline — 10 varied prompts that should fire the sub-agent, each run twice, tracking whether the right agent gets called and passes; (2) turn the sub-agent's system prompt into the body of a new skill.md, rewriting the description to say what it does and when to use it; (3) move the old file aside (don't delete), rerun in a fresh session, and keep the sub-agent only if the folded version scores worse. It's time-consuming to run by hand, but you can have an agent run the analysis for you.
When to keep them
5:19Three cases where sub-agents still earn their place. Context bloat: if research/content flooding in would measurably hurt the main agent, use parallel sub-agents with fresh context windows that summarize findings back. Wrong-tool confusion: if one agent keeps picking the wrong tools, split by context — not by org chart. And the biggest win, parallel research: Anthropic's multi-agent research system beats single-agent Claude Opus by 90% on their internal evals, and sub-agents are great for searching codebases, digging through logs, and read-only reviewers. The trade-off is "a hell of a lot more tokens." But for writing or building, sub-agents aren't always the answer — Cognition's example: one sub-agent built a Flappy Bird background that looked like Super Mario while another built a bird that didn't match it at all. Dependent steps, quick changes, heavy back-and-forth → keep it in the main agent.
Skills as busy work
6:37The same logic extends to skills and CLAUDE.md: every skill description adds to context on every turn whether it's used or not. A Microsoft Research + two-universities paper found skills can create busy work — turning validation checklists and "construction recipes" into mandatory work (a skill saying "always run the full test suite" triggers even on a one-line button change). The cleanup: run /skill doctor (deletes skills that never fire), then read your CLAUDE.md/agents.md and ask "would the current model do this anyway without being told?" — if yes, cut it. Two counterpoints keep this honest: harnesses don't always shrink (Grokbot's coordinator + worker fleet adds guardrails back for non-coding areas models haven't nailed), and a July paper found trimming skills by up to 54% didn't necessarily cut cost per task — skills are small and mostly cached after the first turn. And Anthropic's guide is "a pattern they're seeing from customer work — no pass rates or sample sizes," so test it on your own setup.
Claims checked
The sources are real. Anthropic's "when to fold sub-agents into your main agent loop" (Oct 22), "Building Effective Agents", the multi-agent research system, and Cognition's "Don't Build Multi-Agents" (the Flappy Bird example). The three arXiv papers are in the description.
The "80% system-prompt cut," the "2× faster single loop," and the "90% on internal evals" are the video's reading of Anthropic's guidance — and Anthropic's own guide, per the video, gives no pass rates or sample sizes.
Auto-captions garble "Anthropic" (Antropic/Tropic/Entropic), "CLAUDE.md" (claw.md), "agents.md" (agents.mmd), and "Jediah Katz" (Jadiah Katz). All corrected.