Anthropic recommends consolidating sub-agents into one agent that loads context on demand

Anthropic is recommending that companies evaluate whether they can consolidate sub-agents into a single primary agent that loads context on demand. The company says this design often outperforms multi-agent setups on quality, cost and latency, and it has published a migration kit on GitHub to help teams make the switch, according to a report by The Decoder dated Oct 25, 2026.

What Anthropic found

Anthropic's Applied AI team drew the recommendation from hands-on work across multiple enterprise deployments. It says a single agent using skills frequently beat both monolithic prompt designs and sub-agent setups on quality. In one comparison, Anthropic says the single-agent architecture ran about twice as fast as the multi-agent setup.

The company says newer models, starting with Opus 4.5 and including Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop. That can also reduce costs and wait times.

Two examples illustrate the problem Anthropic describes:

• One company had a dozen specialized sub-agents inside its customer-facing agent. That made a consistent brand experience hard to maintain and hurt usability through high latency.
• In a multi-agent coding workflow, context is lost at every handoff between feature writing, test writing and code review. In the single-agent version, one agent writes both the feature and the tests because it already has the context, while a separate verifier only runs the tests and reports results.

"Conway's Law" for agents

Anthropic says many agent designs follow a familiar pattern: intent detection, a fixed processing pipeline, user segmentation, separate recommendation branches and an orchestrator on top. It argues the model can now handle much of that work itself.

The company calls the organizational version of this problem "Conway's Law" for agents: the architecture becomes a copy of the org chart, with each team shipping its own sub-agent until hundreds of overlapping agents exist with no shared standards. The result, Anthropic says, is a "Frankenprompt," in which teams keep adding contradictory rules to the system prompt.

According to Anthropic, a clean system prompt, four or five lean data tools and fixed output guardrails cover much of the respective use case. It says it cut the Claude Code system prompt for newer models by over 80 percent, partly because skills and tools now load only when needed.

When sub-agents still make sense

Anthropic names three cases:

• A single agent's context window shows measurable performance degradation.
• Multiple tasks need to run in parallel with separate context windows.
• A single agent struggles with tool selection.

The company also warns against merging blindly. In one test with a support agent, folding in a sub-agent reduced latency but broke a previously working process, because the primary agent started answering on its own instead of consulting the right specialist.

Recommended practices

• Package workflows into skills and load tools on demand through a tool search, rather than giving every team its own agent.
• Have a small central team own the system prompt and overall quality.
• Assign owners on each team to maintain skill libraries. Anthropic notes that tool selection tends to degrade beyond 30 to 50 tool definitions.
• Put reference material such as guidelines, product specs and metric definitions in knowledge bases queried through a search tool, or embed it as reference files inside a skill.
• Avoid fetching context through code and injecting it directly into prompts.

The migration kit walks teams through identifying existing patterns, evaluating alternatives and verifying the best approach.

Outside research

The Decoder notes that the findings are consistent with recent research. A study by Vals found that agent teams can be faster than single agents but do not deliver higher quality. According to that study, scaling from ten to 30 or 100 agents brought no quality gains but significantly higher costs. OpenAI developer Eric Provencher has also questioned agent swarms, calling them "almost certainly" wasted money.

What this means

The guidance shifts the default question for agent builders from how to split work across specialists to whether the split is needed at all. Anthropic's stated reasons are concrete: context lost at each handoff, added latency, and system prompts that accumulate conflicting rules. Its own caveat matters as much as the recommendation: the support-agent test shows that consolidation can break working behavior, so the migration kit's evaluate-then-verify workflow is part of the advice, not an optional extra. The performance figures, including the roughly twofold speed difference, are Anthropic's own and come from its deployments.