Multi-agent systems: when one agent isn't enough
When to split work across multiple AI agents instead of one — the patterns that work, the coordination traps to avoid, and how agents hand work to each other safely.
A single AI agent can do a lot. But some work is genuinely a team effort — and forcing it into one agent makes a confused generalist. The question is when to split.
When to use multiple agents
- Distinct roles — a sales agent and a solutions agent have different jobs, tools, and tone.
- Different access — one agent should touch the CRM, another shouldn't; splitting makes least-privilege natural.
- Different owners — separate teams own separate agents, composed across an org.
When not to
If the work is cohesive — one role, one set of tools — a single agent is simpler and easier to reason about. Splitting adds coordination cost; only pay it when the seams are real. (agent orchestration goes deeper.)
The coordination traps
- Shared mutable state. Two agents editing the same record race each other. Put contended writes behind a deterministic operation; keep shared facts in one source of truth.
- Lost context. A delegated agent that doesn't know what its siblings found answers wrong. The fix is to thread the relevant facts into the hand-off, or give it a read of the shared state.
- Runaway delegation. An agent spawning agents spawning agents. Cap it; a leaf worker shouldn't be able to start new work it wasn't meant to.
The pattern that works
One agent owns the objective and delegates focused sub-tasks to others; results come back to it; the whole thing is one durable job with a record of who did what. That's how a sales agent pulls in a solutions agent mid-deal without you wiring anything — see it in action on the sales solution page.
Start with one agent. Add a second only when a responsibility is clearly someone else's job. The enterprise AI agents guide covers the model.
Frequently asked questions
When should I use multiple AI agents instead of one?
Split when responsibilities are genuinely distinct (different roles, tools, or owners) — e.g. a sales agent that delegates a technical question to a solutions agent. Don't split for its own sake; one capable agent is simpler when the work is cohesive.
How do multiple agents coordinate without chaos?
Through durable jobs and explicit hand-offs, not shared mutable state. One agent delegates a sub-task to another, the result comes back, and shared facts live in a single source of truth — not in each agent's head.