Anthropic: When One Agent Loop Beats a Team of Subagents

One Anthropic customer ran about 2x faster after replacing its multi-agent harness with a single agent loop. Another had a dozen specialist subagents and struggled to keep the experience on-brand. For enterprise teams that built routers, planners and domain subagents, Anthropic’s new guidance is a reason to test whether that structure still earns its cost, because the model may now do the job alone.

The Framework

In Anthropic’s article on folding subagents into the main loop, published October 22, the company argues that one loop using progressive disclosure, with skills and tool search loading context only when needed, can beat a multi-agent design on quality, cost and latency. It points to Opus 4.5 as the point where models began handling far more within a single loop.

The organizing idea is that agent architecture shouldn’t follow the org chart. Anthropic warns against agents that become “a copy of their org chart (Conway’s law at work)” and lists the failure modes that follow: details lost in handoffs, overlapping subagents with no clear owners, local evals that skew what leadership sees, and the “Frankenprompt,” where each team adds instructions that conflict and dilute the agent’s voice. The Decoder’s summary notes that Anthropic cut Claude Code’s system prompt by over 80% for more advanced models.

Step 1: Build a Small End-to-End Eval Set

Anthropic says 10 golden examples is enough to start. The point is to measure the whole agent, not each subagent’s subtask, so a merge that helps one team’s score but hurts the user’s outcome shows up.

Step 2: Merge Specialists That Work Together

Candidates are subagents that chain together, touch the same data, or are often called together. Anthropic gives one example: a proposed design with separate intent, reasoning and segment components could often be replaced by one system prompt, four or five thin data tools, and firm output guardrails.

Step 3: Call Remaining Subagents as Tools

Subagents that stay should be called by the main agent as tools, so the main agent keeps the conversation instead of handing it off.

Step 4: Fold In One at a Time

Anthropic suggests turning subagent prompts into skills and API calls into MCP tools, then comparing eval results with and without each change. Tools should be grouped by user intent, since tool choice tends to degrade once a system prompt defines more than 30 to 50 tools. Anthropic’s single-agent loop migration kit chains Claude API skills for building evals, auditing prompts, hill-climbing and cost optimization.

Where It Doesn’t Apply

Anthropic keeps subagents for three cases: a single loop shows measurable degradation from context bloat, tasks run in parallel and benefit from fresh context windows, or one agent struggles to pick the right tools. Its own test shows the risk of merging carelessly. Folding one subagent into another in a sample support agent cut latency but broke a passing process: the main agent answered from the order lookup tool and skipped asking the shipping specialist for the delivery date. The failure appeared in only one of two runs. Stack Archive’s coverage of Vals AI’s agent team test reached a similar view from the cost side: teams cost more and rarely scored higher.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.