Anthropic's Applied AI team now says most subagents should be folded back into… · @willzero · TwKit

@willzero

Anthropic's Applied AI team now says most subagents should be folded back into one main agent loop. One customer's single loop ran roughly 2x faster than its multi-agent harness. The guide came out Oct 22 on Anthropic's site. Across several enterprise deployments, they found one agent with skills frequently beat both the giant single prompt and the subagent design on quality, often at lower cost and latency per task. Their reasoning is simple. Every handoff returns a summary, so detail leaks out at each step. They call it a telephone game. Skills and tools now load on demand, so a long prompt is no longer a reason to split. Claude Code's system prompt for newer models was cut by over 80% this way. Subagents still have three jobs: measurable context bloat, parallel work that needs fresh throwaway context, and an agent that keeps picking the wrong tool or has conflicting instructions. The migration order is concrete. Build an end-to-end eval first, even 10 golden examples. Merge the subagents that always run together. Then turn the rest into skills plus MCP tools, one at a time, and compare scores. Past 30 to 50 tool definitions, tool choice tends to degrade. The best detail is their own failure. Folding one subagent into a sample support agent cut latency but broke a flow that used to pass. The main agent held the order lookup tool, so it answered by itself and never asked the shipping subagent when the package would arrive. The break appeared in just one of two runs, so run every eval more than once. I think the line worth checking against your own system is that many enterprise agents are drawn from the org chart, one team, one subagent. If you run one of those, your call logs will show which pairs always fire together. I'd merge those first.

曝光 136 · 评论 2 · 点赞 1 · 书签 0