Anthropic Told Companies to Stop Building Agent Teams. The Quality Case Is Thinner Than the Cost Case.

On October 22, Anthropic's Applied AI team published engineering guidance aimed at the pattern it helped popularize in the first place: a coordinator agent dispatching work to a roster of specialist subagents. The new advice is to fold most of that back into a single agent loop, where one model repeatedly plans, calls tools, checks results, and revises, pulling in 'skills' and tool definitions only when a task needs them rather than loading everything into the system prompt up front.

Anthropic's own evidence for the switch is concrete on two counts. Claude Code's system prompt for its newer models shrank by more than 80 percent once static subagent definitions were replaced with skills that load on demand. And in one customer migration, cited without naming the company or disclosing comparison conditions, consolidating a multi-agent support system into a single loop cut processing time roughly in half. Anthropic also flagged a practical ceiling: once a single agent's tool list grows past roughly 30 to 50 definitions, models start picking the wrong tool more often, which is one of the few places the guidance still recommends splitting work out.

The independent numbers back the cost story, not the quality one

The clearest confirmation comes from outside Anthropic, though it comes with its own caveat. Vals AI, an independent evals firm, is reported to have run GPT-6 Sol and Claude Opus 5.5 through its Vibe Code Bench — a test that has each model build a complete web app, covering sign-up, posting, search and discovery flows, against 20 to 60 automated checks per spec — once as a lone agent and once as a coordinated team, at both medium and maximum reasoning effort. The team configuration is reported to have burned 1.8 to 5.1 times more tokens than the solo agent in every comparison. On quality, only one of the four solo-versus-team pairings is reported to have shown a statistically significant difference: GPT-6 Sol at medium effort, where the team scored 7.3 points higher. At maximum effort, neither model gained anything measurable from adding teammates. Vals AI's own methodology write-up wasn't independently locatable at publication time, so these figures should be read as a secondhand but consistent report rather than a directly verified result.

Anthropic's broader point — that more agents reliably cost more and rarely help, a narrower and more defensible claim than 'a single loop performs better' — is the one its guidance leans on most. But the company's own public evidence for that narrower claim is thin: beyond the Claude Code system-prompt reduction and the single unnamed customer migration described above, Anthropic has not published its own controlled agent-count scaling results alongside this guidance, which is part of why outside benchmarking like Vals AI's carries unusual weight here.

Anthropic's own quality claim is backed by very little public testing

Anthropic shipped a reference implementation, the single-agent-loop-kit, on GitHub alongside the guidance, along with migration steps: build a small eval set (Anthropic suggests roughly ten examples is enough to start), route subagents through the main agent as callable tools, then convert them into skills one at a time, keeping a subagent only when merging it in measurably lowers eval scores. But the only real-model testing behind the kit on record, as of its publication, was two runs by one person using a single model, claude-opus-5, on 12 tasks; a later expansion to 19 tasks was not run back through the same two-pass comparison. The kit does not accept outside contributions, so any company adopting it has to validate and maintain it on its own evals rather than inherit a tested baseline.

One disclosed test case underlines the risk of migrating on guidance rather than measurement: in a sample conversion, folding a shipping-specialist subagent into the main loop dropped a confirmation step, and the task failed on one of two attempted runs. Anthropic's own framing treats this as expected friction — the whole point of building evals first is to catch exactly that kind of regression before it ships — but it also means the guidance is, by its own account, a direction to test rather than a result to adopt.

The context that makes this worth tracking is that enterprises spent much of 2026 building the coordinator-and-specialists pattern Anthropic itself demonstrated in a widely copied 2025 multi-agent research system writeup, and Anthropic now says those architectures often mirror a company's org chart more than its actual task structure — five-person review teams produce five-subagent pipelines regardless of whether the work benefits from the split. The company pulling back on the pattern it helped spread is notable regardless of how solid the underlying data is. But the two pieces of evidence land very differently: the cost data, replicated by an outside firm across two different frontier models, is hard to argue with. The quality data is still one company's internal anecdotes and one small reference test, which is a thin foundation for the migration a lot of engineering teams are now being told to make.