Callable agents in orchestration

A CallableAgent is a saved agent definition that a workflow node can invoke as a subagent. Each run uses a snapshot taken at run start, so later library edits do not affect it.

agent.rs owns CRUD, validation, and the AgentStore trait. adapters/storage/agent_store.rs persists openflow/agents.json. The orchestration guide covers the rest of the crate.

Data model

engine::graph::CallableAgent is the canonical type. Orchestration re-exports it as AgentDefinition for saved definitions. A run uses an immutable snapshot of the selected records.

Run lifecycle

1. RunCoordinator loads agents through AgentLibrary and freezes the selected records as CallableAgent snapshots.
2. Installed /skill-id tokens anywhere in task prompts resolve through the configured skill catalog. OpenFlow reads the exact SKILL.md files, appends their contents to the node or callable-agent system prompt before the snapshot starts executing, and keeps the resolved paths for later chat turns.
3. The engine uses the snapshot and never re-fetches saved agents during the run.
4. The snapshot is discarded, and the next run takes a fresh one.

Extend agents

1. Add the field to engine::CallableAgent.
2. Cover it in persisted AgentDefinition JSON and compatibility tests.
3. Update agent.rs validation if needed.
4. Update RunCoordinator::start_run() if snapshot semantics change.

References

• agent.rs
• adapters/storage/agent_store.rs
• run/coordinator/mod.rs
• engine::CallableAgent