From the source
Anthropic's Applied AI team says enterprise agents often work better when routers, planners and domain subagents are folded into one main agent loop, and it has published a migration kit to help teams do it.
- Subagents, which have their own context window, tools and instructions under a coordinator, remain useful for verification, parallel work and modular permissioning, but Anthropic says a single agent with skills frequently beat both the one-prompt design and the subagent design on quality, often at lower cost and latency per task.
- The team points to common failure modes in large organizations: context lost in handoffs, hundreds of overlapping subagents with no owner, conflicting rules added to one system prompt, evals optimized only for a subtask, and added latency.
- One customer that compared its multi-agent harness with a single agent loop saw roughly 2x faster runtime with the latter.
- Anthropic recommends progressive disclosure of context through skills and of tools through tool search, so short descriptions load fully only when needed, and says it cut Claude Code's system prompt for more advanced models by over 80%.
- It credits more capable models, starting with Opus 4.5, for handling more in a single loop, and notes thinking cannot be turned off on Opus 5.5 and Fable 5.1.
- Subagents are still advised when a single loop degrades from context bloat, when tasks run in parallel with discardable context, or when tool selection or conflicting instructions become a problem.
