Delegation

Delegation allows an orchestrator agent to assign tasks to specialist agents. The orchestrator coordinates work without executing tasks directly.

How It Works

1. Orchestrator receives a task from the user
2. Orchestrator queries the agent pool (GET /api/agents/pool) to see available specialists
3. Orchestrator delegates via POST /api/agents/delegate
4. Specialist executes the task and returns a response
5. Orchestrator receives the result and continues

Pool Mode vs Chat Mode

Toggle in the chat toolbar:

Mode	Behavior
Pool	Orchestrator delegates via HTTP API to dedicated specialist agents
Chat	Agent uses SDK native subagents (Agent tool) within its own session

Setting Up Delegation

1. Create a workspace with .claude/commands/*.md files defining roles
2. Set one role as orchestrator (workspace settings or agent toggle)
3. Spawn the orchestrator and specialist agents
4. Send a task to the orchestrator — it will delegate automatically

Delegation Rules

The orchestrator receives these rules in its system prompt:

• Idle agent → delegate directly
• Agent not spawned → spawn first, wait 3 seconds, then delegate
• 2+ parallel tasks → use spawn-request (requires user approval)

Delegation Status

Each delegation shows as a card in the orchestrator's chat:

Status	Meaning
queued	Task waiting in specialist's queue
working	Specialist is executing
done	Task completed, response available
error	Task failed

Spawn Requests

When the orchestrator needs multiple specialists simultaneously, it sends a spawn request. A card appears in the chat for the user to approve or deny, with checkboxes for each requested role.

Non-Orchestrator Blocking

Non-orchestrator agents receive a system prompt instruction to execute tasks directly using their own tools, not delegate. The delegation API endpoint returns 403 for non-orchestrators.

Queue Management

Each specialist has a sequential queue. If multiple delegations arrive:

• They execute one at a time (FIFO)
• The stop button drains the queue, rejecting pending tasks
• Results are sent back to the orchestrator as system messages

Pool Context Optimization

The delegation rules (~500 tokens) are injected once in the system prompt. A refresh is sent every 10 turns to prevent context loss after auto-compact.