Anthropic: Building effective agents
Primary guidance on simple workflows, tool use, orchestration, and adding autonomy only when it creates measured value.
← Back to Compute College
Compute College
Single-agent vs multi-agent design
Compare one capable agent with multiple specialized agents and the coordination cost between them.
Plain-English definition
A single-agent design gives one model-driven loop responsibility for a task. A multi-agent design divides work among specialized loops or roles that exchange messages or artifacts. Multiple agents can separate expertise, but coordination adds context, calls, latency, and new failure modes.
Memory trick: More agents mean more coordination, not free intelligence.
Why it matters
Splitting a system does not automatically improve quality. It helps when roles have different tools, permissions, context, or evaluation criteria. Otherwise, one well-bounded workflow may be easier to test and cheaper to run.
- The design makes a hidden tradeoff measurable before it reaches users.
- A clear boundary reduces retries, unsafe actions, or unnecessary capacity.
- The operating choice should be tested against the workload rather than assumed from a demo.
Simple example
A compliance workflow uses one retrieval-and-analysis agent with read-only access, then a separate approval service—not another unconstrained agent—to decide whether a human must review. A multi-agent design is considered only if the roles require genuinely different evidence or permissions.
- The example names the input, decision, and expected boundary.
- A deterministic or human checkpoint owns the part a model should not guess.
- The result can be measured with quality, latency, safety, and cost metrics.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Primary reference
This source explains the method or risk boundary. It does not replace testing the actual model, data, provider, and workload conditions in production.
Source discipline: examples on this page are illustrative; measure the real workload before making a release or capacity decision.
Common mistake
A “manager” agent adds little when ordinary code can delegate the work. Coordination can cost more and make failures harder to attribute.
Practical takeaway
What you can do with this
Compare a single-loop baseline with a proposed multi-agent graph. Measure quality, trajectory length, context transfer, latency, permissions, and cost per accepted result.
- Start with a narrow workflow and explicit failure boundary.
- Instrument the path before optimizing it.
- Review the design whenever model, tool, traffic, or policy changes.
Decision check: what boundary or specialization cannot be achieved with one agent plus deterministic orchestration?
Compute College learning path
AI Engineering
Step 33 of 48: Single-agent vs multi-agent design