Anthropic says one AI agent beats many subagents now that models can handle complexity alone

Anthropic published new architecture guidance yesterday recommending that companies evaluate whether they can replace multi-agent orchestration with a single agent that loads context on demand. In one comparison they ran, the single-agent setup was about twice as fast.

The shift reflects a simple fact: Claude 5.5 can now hold enough context that the original reason for splitting workflows across multiple specialized agents — model context limits — no longer applies. The 200K token window means you can load customer history, product catalog, and current conversation state into one prompt without hitting capacity.

What changed

Early agentic systems split work across subagents because no single model could handle the full scope. You'd route customer support requests to a triage agent, then hand off to a billing agent or a technical agent depending on intent. Each subagent had its own prompt, its own context window, and its own tool access. Orchestration logic managed handoffs.

That pattern made sense when context windows were 8K tokens. It makes less sense now. Claude 5.5 can load the equivalent of a 400-page manual in one call. Most customer support workflows don't need more than that. The orchestration overhead — the latency of routing between agents, the risk of losing context at handoff points, the complexity of managing state across boundaries — often costs more than it saves.

Anthropic's guidance is blunt: if your workflow fits in 200K tokens, try collapsing it into one agent first. Use prompt caching to keep retrieval fast. Load the full context on every call instead of fragmenting it across subagents.

The speed difference

In the architecture comparison Anthropic shared, the single-agent setup ran about twice as fast. The multi-agent version spent time on handoffs, re-authenticating tool access, and re-loading overlapping context. The single-agent version loaded everything once and executed the workflow in one pass.

That's not a small difference. For a voice agent on a real phone line, 2× faster execution means the difference between a customer waiting through dead air and a conversation that feels responsive. For a high-volume workflow like prior auth or claims processing, 2× throughput means half the compute cost.

The trade-off is prompt complexity. A single-agent prompt that handles triage, billing, technical support, and escalation is longer and harder to debug than four focused prompts. But Anthropic is betting that the tooling for managing one complex prompt — version control, diff tracking, A/B testing — is more mature than the tooling for orchestrating four agents.

The migration kit

Anthropic published a migration kit on GitHub to help teams move from multi-agent to single-agent setups. It includes:

• A prompt template that consolidates multiple subagent prompts into one structured block
• A context-loading pattern that uses prompt caching to keep retrieval under 500ms
• A tool-access policy that grants the single agent the union of all subagent permissions
• A fallback pattern that re-routes to a human if the single agent can't resolve in one pass

The kit assumes you're running on Claude 5.5 or later. It won't work on earlier models with smaller context windows. It also assumes your workflow can be represented as a single decision tree, not as a branching state machine that requires parallel execution.

When multi-agent still makes sense

Single-agent setups don't work for every workflow. If you're running a long-running process that spans multiple user sessions — like a multi-day sales cycle or a multi-step approval chain — you still need orchestration to manage state persistence. If you're running workflows that require parallel execution — like processing 10 support tickets simultaneously — you still need multiple agent instances.

But for synchronous, single-session workflows — the kind we deploy for Best Limo, Bergsify, and DreamersJoy — the single-agent pattern is now the default. It's faster, cheaper, and easier to debug. The orchestration complexity isn't worth the modularity benefit anymore.

Anthroplic's guidance is the first time a frontier lab has explicitly recommended moving away from multi-agent architectures. It's a sign that context windows have crossed a threshold where the old splitting logic no longer holds. The models are big enough now that you can fit the whole workflow in one call.