Anthropic's Applied AI team says stop copying org charts with agents: eval first, fold subagents in one at a time
ombre-ai · reddit · 2026-10-29
Anthropic's Applied AI team (Oct 22) published enterprise agent guidance that reads like a post-mortem on multi-agent builds: architectures that mirror the org chart (Conway's law) suffer context loss at handoffs, hundreds of ownerless subagents, contradictory "Frankenprompts", fragmented evals, and latency costs — one customer got 2x faster runtime after moving to a single loop.
Their rule: split by context, not problem type. A subagent should own work needing little from the rest of the system and return a compact result.
The migration method: build a small end-to-end eval first (10 golden examples suffice), fold in one subagent at a time (prompt → skill, API calls → tools), keep it only if the folded version scores worse, and run each task more than once — a regression in their own support-agent test appeared in just one of two runs. Subagents are still recommended for parallel work, measurable context bloat, or tool-selection trouble.
More from coding & agent
- MemPalace hits 49K stars: local-first AI memory with 96.6% LongMemEval recall, zero API calls — thisguyknowsai · 2026-10-29
- GitHub ships hands-on cookbooks for Copilot agentic features post-Universe — marlene_zw · 2026-10-29
- Lean 4-person AI startup Physera hires to build training and eval infra for multimodal coding agents — himanshustwts · 2026-10-29
- Using a Claude subscription inside third-party apps runs into ToS limits — ZookeepergameOdd6078 · 2026-10-29
- Matt Pocock fixes Codex subagents inheriting parent context instead of starting fresh — mattpocockuk · 2026-10-29
- Indie dev builds an LLM cost guard that alerts on unusual API usage — WittyBoysenberry9860 · 2026-10-29