Use case · Updated 2026-06-02
Multi-agent orchestration
Fork child harnesses with scoped tools and prompts, wait for results, run critics on outputs, and share memory on the session bus — orchestration tools are harness-scoped and recreated per child, not copied from parent.
spawn_agent and children
forkChild() creates a new AgentHarness with its own registry slice. ORCHESTRATION_TOOL_NAMES stay on the parent — children cannot spawn workflows or refresh world context at root. finalizeChildSpawnTools wires share_agent_context, read_agent_context, and dependency injection: dependsOn agents auto-receive upstream results via buildUpstreamDependencyContext.
Verification and debate
- verify_result — run when final answers are code- or path-heavy (AGENT_CRITIC).
- evidence_critic, path_critic, policy_critic — structured challenges.
- reflect_debate — adversarial pass between positions.
- verify_contract — check execution contracts from decompose_goal.
Contract-driven tool activation
contract_tool_mapper maps execution contracts to tool families so children start with the right capabilities (web, code_intel, memory) without loading everything. Intent inference and spawn_tool_inference further narrow provisioning.
Telemetry
Subtask telemetry surfaces in the web SubtaskInspectorModal and session JSONL. cancel_agent stops runaway children. list_agents shows the fleet state mid-turn.