Anthropic 的单代理工作流建议

Anthropic's applied AI team spent the last year watching enterprise customers build agent systems that look like org charts: a router on top, a planner underneath, one specialist subagent per function. Their new guidance, out this week, says most of that should come apart. One customer ran roughly twice as fast after collapsing their setup into a single agent with a shared context window. The failure mode Anthropic names is specific: every subagent handoff returns a summary, and every summary drops detail. They call it a telephone game. The replacement pattern is one model, one context window, an open ended tool loop with turn and cost budgets enforced by the harness rather than the prompt. Each subagent's job becomes a skill. Its API calls become MCP tools. Approval gates move into the tools themselves. None of this works without progressive disclosure, loading instructions only when a task actually needs them. Anthropic says that technique alone cut Claude Code's system prompt by more than 80% on their newer models. It's also why the pattern couldn't have shipped two years ago: the models weren't reliable enough to manage that much ambiguity in one context window. Worth noting this isn't a totally new argument. Cognition built Devin on a single sequential agent back in 2024 explicitly because they'd watched multi-agent swarms produce confident, coordinated nonsense. LangGraph and CrewAI still default most users toward multi-agent graphs. Anthropic is now telling its own enterprise customers the opposite of what half the agent tooling industry sells by default. The honest caveat in their own writeup: folding one subagent lowered latency but broke a process in one of two test runs, because the agent skipped a shipping check it used to delegate out. Their fix isn't "don't fold," it's "fold one at a time and eval every change with real examples." That's a less exciting takeaway than the speed number, and probably the more important one. https://x.com/Arindam_1729/status/2114696945772249550 Arindam Majumder 𝕏: Anthropic just told teams to delete most of their subagents! (one customer ran ~2x faster after moving to a single agent loop) for a year, the default agent design looked like an org chart. a router on top, a planner, then one specialist subagent per team. anthropic's applied