Back to Glossary
Multi-Agent Coordination
Synonyms: multi-agent orchestration, agent orchestration, orchestrator-worker pattern, supervisor agents, agent handoffs
Do not index
Definition
Multi-agent coordination is how several AI agents split a task between them and merge their work into one result. Usually a lead agent (the orchestrator) breaks the job into pieces and hands each one to a specialist agent. For designers, the job is showing users which agent is doing what, and giving them human-in-the-loop controls to step in.
Use cases
When coordination breaks, users get slow or contradictory results and can't tell which agent caused the problem.
- Agents that disagree: A contract tool runs a pricing agent and a legal agent on the same clause. One calls it fine, the other flags it. Both answers get equal weight, and the user picks the wrong one.
- The 90-second spinner: A trip planner runs 4 agents behind one loading icon. Users think it froze, refresh, and kill the run.
- Lost context: Cognition split a Flappy Bird clone between 2 subagents. One built a Super Mario-style background. The other built a bird that didn't match it.
How it's used in practice
- Name each agent in the progress view: "Flights agent: checking 3 airlines" lets users follow each handoff.
- Show partial results as they land: Waiting for the final summary makes a fast system feel slow.
- Let one agent talk to the user: The lead agent owns the conversation and merges results, so users hear 1 voice.
- Show conflicts on purpose: Put both positions side by side with sources and let the user decide.
- Review the plan before agents start: Let users edit the task breakdown first. Microsoft's Magentic-UI calls this co-planning.
Ask engineering early: do the agents only read, or also write? Agents that search and summarize rarely clash. Agents that edit files or book things make conflicting decisions, so budget for approval steps and undo.
Challenges & limitations
- Higher cost and longer wait times: Multi-agent systems can consume significantly more resources than a single-agent setup, increasing costs and response times.
- Difficult to debug: Coordinating multiple agents makes it harder to identify failures, resolve conflicting outputs, and understand where things went wrong.
- Not suitable for every task: Tasks that require all agents to share the same context or work closely on the same problem may not benefit from splitting the work across multiple agents.
Free resources
- Designing For Agentic AI: Practical UX Patterns (Smashing Magazine) — Victor Yocco's 6 patterns for control, from intent previews to undo, with metrics for each.
- Magentic-UI, an experimental human-centered web agent (Microsoft Research) — how a multi-agent team supports co-planning, action guards, and plan learning.
- How we built our multi-agent research system (Anthropic) — the orchestrator-worker setup behind Claude's Research feature, including token costs and failure cases.
