Core multi-agent orchestration patterns
A supervisor pattern uses one coordinator to assign work to specialist agents. A pipeline pattern passes artifacts through sequential stages. A parallel-review pattern runs independent agents at the same time and then reconciles their outputs.
More agents are not automatically better. The architecture should match the dependency graph of the work and use the smallest number of agents that creates a measurable quality or throughput advantage.
- Supervisor and specialist agents.
- Sequential pipeline or handoff.
- Parallel research or implementation branches.
- Reviewer and critic stages.
- Independent final verification.
State, recovery and duplicate-work prevention
Long-running orchestration should persist mission state outside the worker process. Each task needs a durable status, ownership or lease information, retry history and a record of completed outputs.
Without durable state, worker restarts can duplicate expensive model calls or lose completed work. Recovery should resume from the last verified checkpoint instead of restarting the entire mission.
How gateways support multi-agent systems
Agent frameworks decide what work happens next. The gateway underneath them can centralize provider credentials, routing, budgets, rate limits, fallback and telemetry for every model call the agents make.
Keeping those layers separate makes it easier to change an agent framework without rewriting provider infrastructure, or to use the same gateway for both agentic and non-agentic applications.