Anthropic Tells Teams to Fold Subagents Into One Agent Loop
Anthropic’s own guidance now says many enterprise agents built from several subagents would perform better as a single agent loop that loads skills and tools on demand. In its best-practices guide, Anthropic reports that across several enterprise deployments a single agent with skills often beat both one-prompt and subagent designs on quality, usually at lower cost and latency. One customer saw a single loop run roughly twice as fast as its multi-agent harness.
The failure modes will be familiar to anyone who has shipped a multi-agent system: context lost across handoffs, subagent sprawl, conflicting rules in a patched-together system prompt, evals that test only one fragment, and added latency. Anthropic says it cut Claude Code’s system prompt by over 80 percent for more capable models. Subagents still earn their place when context bloat measurably hurts a single loop, when parallel work needs fresh context windows, or when one agent struggles to pick tools. Tool selection tends to degrade beyond 30 to 50 definitions, so Anthropic advises tool search.
The migration advice is the useful part, because it starts with outcomes rather than architecture. Build an end-to-end eval set of about ten golden examples. Merge tightly coupled specialists first. Fold subagents in one at a time, compare scores after each change, and run each task more than once. A GitHub migration kit accompanies the post.
That eval set doubles as your production yardstick. Anthropic’s separate OpenTelemetry trace export sends every agent turn, model call and tool use to the monitoring tools a team already runs, and Ramp reports 10 to 20 percent lower token usage after scoring its agent traces in Langfuse.
If your agent has more than a couple of subagents, count the handoffs, write the ten examples, and test whether one loop ships the same outcome with less machinery. Simplifying is a result you can measure.