THE SHORT ANSWER

Use one agent when a single context and tool set can handle the task clearly. Consider multiple agents when work separates into genuinely distinct specialties or independent parallel tasks. Account for coordinator decisions, handoffs, duplicated context, conflicting outputs, extra calls and propagated errors.

One agent is a strong default

A single agent keeps the goal, state and trace in one place. It avoids handoff protocols and reduces duplicated context. If one evaluated model can select the necessary tools and remain within a clear loop, extra roles may only rename additional model calls.

Improve the single system first: clearer tools, selected context, deterministic checks and routing between easy and difficult cases. Multi-agent design should solve a measured limitation.

Recognise distinct multi-agent patterns

Patterns and their trade-offs
PatternPotential valueAdded risk
SpecialistsDifferent instructions or tools for distinct domainsWrong routing and fragmented context
Coordinator and workersDynamic decomposition of unpredictable subtasksCoordinator error and expensive synthesis
Parallel reviewersIndependent checks or perspectivesDuplicated cost and false confidence from correlated errors
Sequential handoffsClear responsibility by stageInformation loss and unclear ownership

Evidence & context: Anthropic

Treat handoffs as interfaces

A handoff should state the subgoal, relevant evidence, decisions already made, allowed tools, output contract and stop condition. Passing an entire transcript can inflate cost and transfer irrelevant or unsafe instructions. Passing only a summary can omit crucial evidence.

The coordinator should verify worker outputs before combining them. One worker's unsupported assumption can otherwise become another worker's premise and compound across the system.

Evidence & context: Anthropic

Require evidence for added agents

  • Compare against a strong single-agent baseline.
  • Measure accepted outcomes, total calls, tokens, latency and review.
  • Score routing and handoffs separately from final answers.
  • Test partial failures and conflicting worker results.
  • Keep ownership and escalation visible across roles.

Evidence & context: Microsoft Learn

Sources & further reading

  1. Building effective agents

    Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.

  2. Demystifying evals for AI agents

    Anthropic. A provider's engineering guidance on multi-turn agent evaluation, checked 13 September 2026. Examples inform evaluation design but do not establish universal pass thresholds.

  3. Evaluate a model router

    Microsoft Learn. Official guidance for evaluating routing across representative workloads using quality, cost, latency and policy criteria. It is not evidence that routing always improves results.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗