Agent topology

Topology is the shape of an agentic system: how many agents exist, what each is responsible for, how work passes between them, and where a run can be stopped.

The count is a cost decision

Every additional agent adds a handoff, and every handoff is a place where context is lost, an error is passed on as fact, or latency accumulates. The right number is the smallest one that produces a measurable benefit. A large share of workflows described as multi-agent would run better as one agent with well-chosen tools, or as a deterministic sequence with a model at a single step.

When more than one is justified

Three conditions, and at least one must genuinely hold. Specialization: a step needs a different model, a different tool set, or a different set of permissions. Parallelism: independent branches can run at once and wall-clock time matters. Context isolation: one agent should not see what another handles, whether for quality or for permission reasons.

Common shapes

Sequential, for known steps. Supervisory, where a coordinator delegates and evaluates. Event-driven, where work is triggered by state changes rather than by a caller. The choice follows the problem; none of them is a house style.

What topology cannot fix

A topology change will not repair a step that fails on its own. Splitting a failing agent into three usually produces three failing agents and two new handoffs. Fix the step first, then decide whether the shape needs to change.