Multi-agent systems: coordination, cost and acceptance

When one agent prepares data and another drafts a report, their intermediate handoff needs an agreed structure. Include sources, units, reporting period and unresolved fields. If the first returns a number without context, the second…

Ian · 2026-05-19

Practical guide

A multi-agent system divides tasks and combines results. It can help with independent subtasks, although coordination adds complexity. Compare it with one assistant or a defined workflow before adoption.

Planning a first application

  1. Bound functions, sources and permissions per agent.
  2. Define result sharing and conflict handling.
  3. Keep a person responsible for accepting the complete result.

What to check before deciding

Measure correctly completed tasks, retries, cost and time. Logging should support reconstruction of actions and failures as implemented. Do not attribute improved accuracy to additional agents without a reproducible comparison.

Define a contract between agents

When one agent prepares data and another drafts a report, their intermediate handoff needs an agreed structure. Include sources, units, reporting period and unresolved fields. If the first returns a number without context, the second may present it with unjustified certainty. The coordinator should validate that contract before proceeding, just as it would validate a conventional tool response.

Control repetition and disagreements

Test an incomplete handoff, conflicting results and an agent that does not respond. Decide when to retry, switch alternatives or stop work. Allocate an overall budget and step limits so that a circular conversation cannot consume all resources. Each retry should retain the task identifier to prevent duplicate external actions.

Evaluate the final report against a simpler single-assistant workflow. Record coordination, review time and errors propagated between stages too. A system can produce many responses without completing its assignment. Acceptance should verify that supporting sources reach the final document and uncertainties survive consolidation. Broader autonomy should follow observed outcomes, not the number of participants. Clear stopping conditions are particularly important when no participant has enough evidence to resolve a disagreement.

Primary reference: Anthropic · Building effective agents

Talk to Innovatec in a free 30-minute introductory conversation to assess your needs.