Anthropic: Folding Subagents Into a Single Agent Loop Often Wins on Quality, Cost and Latency

Claude (Anthropic) · rss · 2026-10-22

Anthropic's Applied AI team reports that across enterprise deployments, a single agent with skills frequently beats both one-prompt designs and orchestrator-subagent architectures on quality, often at lower cost and latency (one customer saw 2x faster runtime). Key points:

Failure modes of multi-agent orgs: context loss in handoffs, subagent sprawl, conflicting "Frankenprompts", watered-down localized evals, and architectures mirroring org charts (Conway's law).
Why single loops work now: shared context yields richer outputs; progressive disclosure of skills/tools prevents context bloat (Claude Code's system prompt cut by 80%+); models since Opus 4.5 handle far more per loop, narrate progress and self-check, making some harness layers redundant.
Subagents still fit when context bloat degrades performance, parallel fresh-context tasks are needed, or tool selection/instructions conflict.
Recommendations: one small central team owns quality and the system prompt; if going multi-agent, start with the simplest pattern and split by context boundaries, not problem type, to avoid the telephone game.

Includes a migration kit to identify foldable subagents and verify alternatives.

Original post →

Anthropic: Folding Subagents Into a Single Agent Loop Often Wins on Quality, Cost and Latency

More from coding & agent