🗓️ 19092026 2037
In the agent-as-tool pattern, one primary agent owns the goal and state. It calls specialized agents like tools, evaluates each result, then chooses the next step.
primary agent -> specialist -> result -> primary agent decides next action
When it helps
Use it when a result changes the plan. For example, a debugging agent can ask one specialist to inspect logs, then choose a targeted test based on that answer. The primary agent retains the context that connects those decisions.
Difference from a coordinator
A coordinator_router_pattern usually delegates independent tasks and combines their answers. Agent-as-tool is more sequential: each specialist answer can redirect the primary agent. That makes it useful for investigation, but the primary agent can become a context and latency bottleneck.
Use a single agent first. Add this pattern only when separating a narrow expert task makes the primary agent's decision easier.