Claude's New Guidance: Ditch Subagents for a Single Agent Loop
Multi-agent orchestration is the default in many enterprise stacks, but Anthropic's measurements show it doubles latency and loses detail to lossy summaries. This guidance hands teams a practical migration path: fold subagents into the main loop, move business logic into Skills and tools, and let one model hold the whole context.
Anthropic's Applied AI team has been measuring enterprise deployments and now says the common multi-agent pattern, an orchestrator coordinating domain-specific subagents, mostly fails in production. A single agent with a few Skills and four or five lightweight tools, protected by guardrails, beats both universal-prompt and subagent schemes on quality while cutting cost and latency. Their customers measured the single Agent + Loop architecture at roughly twice the speed of multi-agent frameworks.
The mechanism behind this is progressive disclosure: agents see only short descriptions of Skills and tools up front, and load full definitions through Tool Search only when needed. That removes the main reason teams split work across subagents, the context limit, and lets one model keep a coherent global context across the whole task. Claude Code has already cut system prompts for stronger models by over 80% using this approach.
Migration involves merging domain subagents into the main loop, converting prompts into Skills, API calls into MCP tools, and evaluating each merge with end-to-end tests. Ownership moves from agents to Skills and tools, with a core team guarding system prompt quality. Anthropic names the failure modes it keeps seeing: subagent proliferation, patchwork prompts, local unreviewed evals, and lossy summaries between agents, which it diagnoses as Conway's Law and a telephone game.
The advice to move decisions from harness to model implies the orchestration layer will thin out as model capacity grows, but guardrails still have to live somewhere, and Anthropic puts them inside tools rather than in standalone agents.
Conway's Law critique is the sharpest practical point: when agent boundaries mirror team boundaries, information loss is not a bug in the model but a structural consequence of the architecture.
Anthropic's own toolkit promotion muddies the message, but the underlying claim, that a single loop with progressive disclosure removes the main reason for subagents, is testable with their own eval approach.