Anthropic Urges Firms to Fold Subagents into One Agent Loop
5 min read
Anthropic is recommending that companies evaluate whether they can consolidate sub-agents into a single agent that loads context on demand, marking a significant shift in how enterprise AI systems should be architected. The company published guidance on October 22, 2026 titled “Building agents: When to fold subagents into your main agent loop,” based on patterns its Applied AI team has observed across multiple enterprise deployments.
The company found that a single agent with on-demand context often outperforms multi-agent setups in quality, cost, and latency. One customer that evaluated its multi-agent harness against a single agent loop saw roughly 2x faster runtime with the single loop.
Why Multi-Agent Systems Fall Short
The company’s experience shows that many specialized sub-agents can lead to higher latency, rising costs, and information loss at every handoff. Anthropic warns that giving each team or problem type its own subagent turns the architecture into a copy of the org chart, citing Conway’s law, with agents playing what it calls a “telephone game,” passing information back and forth while each handoff degrades fidelity.
Across several enterprise deployments, “a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task,” according to Anthropic’s Applied AI team.
A Vals AI benchmark found agent teams cost 1.8 to 5.1 times as much as single agents, and only one score difference was statistically significant. Newer models starting with Opus 4.5, along with Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop, which can also reduce costs and wait times.
When Subagents Still Make Sense
Despite the push toward consolidation, Anthropic acknowledges that subagents remain useful in specific scenarios. The guide is clear that subagents still have a place when a single agent hits measurable performance loss from context bloat, or when several tasks can run in parallel with fresh context windows you can throw away afterwards.
Anthropic keeps subagents for three cases: measurable context bloat, parallel work in fresh contexts, and trouble choosing tools, with its rule being to divide by context, not by problem type.
In a separate test, a multi-agent Claude workflow found 66 of 70 planted bugs, versus at most 27 for a single agent, demonstrating scenarios where multiple agents can provide value for complex debugging tasks.
Migration Tools and Enterprise Impact
Anthropic has published a migration kit on GitHub to help teams make the transition. While Anthropic has released an open-source migration tool to help firms analyze and consolidate their architectures, it warned that merging subagents can occasionally cause task failures and requires rigorous end-to-end evaluation.
Skills and tool search use progressive disclosure: the agent sees short descriptions and loads the full content only when needed, which removes a main reason teams split work into subagents. The guide recommends grouping tools by user intent, because past roughly 30-50 tool definitions, tool choice tends to degrade.
When Anthropic reviews proposed designs, it often sees an intent classifier, a fixed multi-step reasoning pipeline, user-segment buckets, separate recommendation branches and a multi-agent orchestrator stacked together, but in those reviews, it says, one clear system prompt, four or five thin data tools and firm output guardrails would cover most of the use case.
Broader Implications for Agent Design
The guidance represents a fundamental rethinking of how enterprise AI agents should be structured as models become more capable. Both sides agree on the same starting point though: don’t hand-build a fixed team of agents for every task, and let one agent act as an orchestrator and only split into sub-agents when the job actually needs it.
The guide is dated October 22, 2026, on when to fold subagents into one agent loop with skills and tools loaded as needed, and when to keep them. Anthropic published a guide telling teams that build agents to fold most of their subagents back into one agent loop, and it has spread fast among agent builders, though the version going around is that subagents are finished.
Key Facts
- Publication date: Anthropic released the guidance on October 22, 2026
- Performance improvement: One customer saw roughly 2x faster execution with a single-agent loop
- Cost differential: Agent teams cost 1.8 to 5.1 times as much as single agents in benchmark testing
- Model support: Opus 4.5, Opus 5.5, and Fable 5.1 can manage complex plans in a single loop
- Tool limits: Past roughly 30-50 tool definitions, tool choice performance tends to degrade
- Migration support: Open-source migration kit available on GitHub
Sources
- Anthropic says one AI agent beats many subagents now that models can handle complexity alone
- Anthropic Says Fold Subagents Into One Agent Loop: How to Do It
- Multi-Agent Teams in Claude Code and Codex: Worth It or Not?
- One Agent or Subagents? What We Learned Running AI Employees
Sources
- Anthropic says one AI agent beats many subagents now that models can handle complexity alone
- Anthropic Says Fold Subagents Into One Agent Loop: How to Do It
- claude subagents: When to Use One Agent vs Many (A Practical Setup Guide) – DEV Community
- AGTP on X: “Anthropic just told companies to use fewer subagents in their AI systems today. Here’s what you need to know. Anthropic’s team is now advising companies to remove unnecessary subagents. Some expert agents are being merged back into the main agent loop. Across multiple enterprise projec… / X
- One Agent or Subagents? What We Learned Running AI Employees
- Anthropic Advises Companies to Cut AI Subagents for Single-Agent Architectures
- Multi-Agent Teams in Claude Code and Codex: Worth It or Not?
