Anthropic now says fold most subagents into one loop with skills — multi-agent often costs more for less

Why this label: Anthropic's own Applied AI guidance and figures. We re-read the article and the THE DECODER write-up on 2026-10-28; we ran no evals ourselves. Distinct from our earlier cards on one developer's 394-subagent lockout and Max-plan burn — this is the vendor's enterprise recommendation.

Tried

Ship an enterprise or product agent as an orchestrator plus a specialist subagent per team or problem type (the org-chart pattern Anthropic calls out).

Broke

Handoffs lose context, latency stacks, and cost rises. Anthropic reports comparisons where a single agent loop ran about 2x faster than the multi-agent harness, and where a dozen specialist subagents made brand voice and latency worse. They also cut Claude Code's system prompt for newer models by over 80% partly by loading skills and tools only when needed.

Cause

• confirmed Anthropic's own article (Building agents: When to fold subagents into your main agent loop) states that a single agent with skills frequently outperformed subagent designs on quality, often at lower cost and latency, across several enterprise deployments.
• confirmed They list when subagents still win: measurable context-window degradation, parallel throwaway work with fresh windows, or tool-selection / conflicting-instruction failures in one loop.
• confirmed THE DECODER summary (2026-10-25) restates the ~2x runtime comparison and the GitHub migration kit; we treat the article itself as the primary source.

“Confirmed” means the source states it as established; “suspected” means the source treats it as a theory or it is unsettled. We did not verify either ourselves.

Fix

Anthropic's suggested end state: one model, one context window, open-ended tool loop with turn/cost budgets; turn domain prompts into skills and API calls into tools; call remaining subagents as tools rather than handing off the conversation. Use their migration kit to build a small end-to-end eval, fold one subagent at a time, and keep it only if the folded version scores worse. Watch for regressions — in one support-agent test, folding cut latency but broke a process that needed a specialist.

Prevention

Before adding another specialist subagent, ask whether a skill plus tools in the main loop would do. Measure task completion, cost, and latency on a shared eval — do not copy the org chart into the agent graph by default.

Sources

• Anthropic — Building agents: When to fold subagents into your main agent loop
• THE DECODER summary (2026-10-25)

← All stories