Sub-agents

Spawn child harnesses that share the same backend and user-scoped memories. Task-scoped work stays private to the child.

spawn()

const child = harness.spawn({
  agentId: "researcher",
  sessionId: "research-task-1",
})

The child harness:

• Shares the same backend connection
• Inherits the parent's userId
• Has its own agentId and sessionId
• Can read the parent's user-scoped and agent-scoped memories
• Can write task-scoped memories that stay private

Memory visibility

// parent writes a user-scoped memory
await harness.memory().write({
  content: "User prefers concise answers",
  scope: "user",
  embedding,
})

// child can read it
const results = await child.memory().search({
  embedding: queryVec,
  scope: ["user"],
})
// results include "User prefers concise answers"

// child writes a task-scoped memory — private
await child.memory().write({
  content: "API uses REST, not GraphQL",
  scope: "task",
  embedding,
})

// parent cannot see task-scoped memories from child

Closing

child.close()

Closing a child harness does not close the shared backend. Only the root harness's close() method closes the backend connection.

Nested spawning

Children can spawn their own children. The same rules apply — shared backend, inherited userId, isolated task scope.

const grandchild = child.spawn({
  agentId: "sub-researcher",
  sessionId: "sub-task-1",
})