Anvia

Agent as a tool ​

agent.asTool() creates a normal Anvia tool whose input is a focused prompt and whose output is the child agent's final text.

1. Build a narrow specialist ​

ts
const policyAgent = new Agent({
  id: 'policy-review',
  name: 'Policy reviewer',
  model: policyModel,
  instructions: [
    'Review the supplied draft for policy risk.',
    'Return findings and recommended changes.',
    'Do not write the final customer response.',
  ].join('\n'),
  maxTurns: 2,
})

Give the child only the tools and context required for its role. It may use a different model from the coordinator.

2. Expose it with a precise description ​

ts
const policyReview = policyAgent.asTool({
  name: 'policy_review',
  description: 'Review a draft customer response for policy risk.',
  maxTurns: 2,
  suspension: 'reject',
  stream: true,
})

name is the stable tool name. description tells the coordinator when to delegate. maxTurns overrides the child run limit for this tool call. suspension: 'reject' is required; agent tools cannot suspend. stream: true forwards child runtime events through a streamed parent run.

Without stream: true, the parent still receives the final child output as its tool result.

3. Add it to a coordinator ​

ts
const coordinator = new Agent({
  id: 'support',
  model: coordinatorModel,
  instructions: [
    'Use policy_review before sending a high-risk answer.',
    'Resolve the findings and write one final response yourself.',
  ].join('\n'),
  maxTurns: 6,
  tools: [policyReview],
})

Keep specialist names and descriptions distinct. Overlapping tools make model routing ambiguous.

4. Pass a self-contained task ​

asTool() calls the child with a stateless prompt. It does not inherit the parent transcript, memory session, user ID, tenant metadata, or retrieval context.

Tell the coordinator to include only the facts and draft the specialist needs. Use the explicit session wrapper in Memory boundaries only when the child truly needs durable continuity.

5. Avoid approval inside an agent tool ​

An agent used through asTool() requires suspension: 'reject'. If the child reaches an interaction outcome, Anvia cancels that child phase and reports an agent-tool error.

Run interaction-capable work directly at an application boundary, or expose the side effect as a protected parent tool whose interaction the product can continue.

Next, consume child events.

Built for Anvia.