--title=Actor model: Integration with agentd

Integrate actor model with existing agentd infrastructure.

Overview

This is Phase 5 of the actor-native agent orchestration system. Connect the actor primitives to existing agentd components: workflows, events, watch, and CLI.

Requirements

Workflow YAML Compatibility

Existing workflow YAML continues to work. It becomes sugar for actor operations:

steps:
  - name: research
    run: research.md
  - name: implement
    run: implement.md

Equivalent to a root actor that: 1. Creates child for research, waits for result 2. Creates child for implement with research context, waits for result 3. Sends final result to customer

CLI Changes

# Run a single task (spawns root actor)
agentd run task.md

# Run a workflow (spawns root actor that follows workflow)
agentd run workflow.yaml

# Watch now shows actor graph
agentd watch

# Watch specific actor and descendants
agentd watch actor-id

Event Integration

New event types for actor operations:

• actor_create - actor spawned child (parent_id, child_id, capabilities)
• actor_send - actor sent message (from, to, customer, payload summary)
• actor_receive - actor received message (actor_id, message_id)
• actor_become - actor changed behavior (actor_id, new_behavior summary)
• actor_complete - actor finished (actor_id, result summary)

Events go to existing _/events/{run_id}.jsonl system.

Watch Integration

• Show actor tree, not just flat runs
• Display message flow between actors
• Show capability budgets (cost remaining, children remaining)
• Termination: watch exits when actor graph fully drained

System Prompt

Every actor receives actor-aware system prompt additions:

• Your actor_id
• Your parent_id (if not root)
• Your customer_id
• Your capabilities summary
• Instructions for using send, create, become

Entry Point

• Human runs agentd with a task
• System creates root actor with full capabilities
• Root receives initial message with task
• Computation proceeds via message passing
• System terminates when root completes and graph drains

Implementation Notes

Files to modify

• Omni/Agentd.hs - CLI, workflow->actor translation, root actor creation
• Omni/Agent/Events.hs - New event types
• Omni/Agent/Watch.hs - Actor graph display
• Omni/Agent/Prompts/ - Actor system prompt additions

Testing

• Test workflow YAML still works
• Test single task.md works
• Test events capture actor operations
• Test watch shows actor graph
• Test termination detection

Dependencies

• Task: Actor model: Message passing infrastructure
• Task: Actor model: CREATE primitive
• Task: Actor model: BECOME primitive
• Task: Actor model: Capability enforcement

References

• Spec: _/llm/actors.md (Integration section, Compatibility section)
• Current workflow: Omni/Agentd.hs
• Events: Omni/Agent/Events.hs