← sources

Anthropic's own Applied AI team says too many enterprise agent designs just copy the org chart. Their fix is an eval first, then folding subagents in one at a time. What made you keep yours?

source: reddit (reddit:AI_Agents)published 2026-10-29 09:07ingested 2026-10-30 02:36

retained text

Anthropic's Applied AI team posted guidance on Oct 22 that reads like a post-mortem on a lot of multi-agent builds. Their observation is that too often every team or problem type gets its own subagent, so the agent architecture ends up copying the org chart (they cite Conway's law).

The failure modes they list will sound familiar to anyone who has worked in a big org. Context gets lost in handoffs. Hundreds of overlapping subagents pile up with no owner. Every team adds a sentence to the system prompt until the rules contradict each other, which they call the Frankenprompt. Each team writes evals tuned to its own subtask, so nobody owns whole-system quality. And the extra layers cost latency: one customer measured a roughly 2x faster runtime after moving to a single loop, though Anthropic does not name the customer.

Their suggested rule is to split by context instead of by problem type. A subagent should get a piece of work that needs little from the rest of the system and hand back a compact result. Split by problem type and each handoff degrades the information, which they describe as a telephone game.

The migration method is the part I would steal. Build a small end-to-end eval first (they say even 10 golden examples is enough to start), fold in one subagent at a time by turning its prompt into a skill and its API calls into tools, and keep it as a subagent only if the folded version scores worse. They also say to run each task more than once, since a regression in their own support-agent test showed up in only one of two runs.

The comparisons are Anthropic's own, from enterprise deployments, and the underlying data is not published. They still recommend subagents for parallel work, measurable context bloat, or tool-selection trouble.

If you run a multi-agent setup, what made you keep a subagent when you tried folding it in, or what broke when you did? And does one team own your evals, or does each team write its own?

document fetched 2026-10-30 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xiew6a/anthropics_own_applied_ai_team_says_too_many