3 · Two protocols

Public client surface vs private worker surface — different jobs, different shapes.

Agent Protocol (client → plane)

Public HTTP API (plus SSE / WebSocket for streaming). Same conceptual surface many Agent Protocol clients already expect: threads, runs, stream, cancel, resume, store. It never exposes reclaim generations, runner kinds, or queue internals.

• HTTP — create/get/search, store access
• SSE — one-way token/event stream to the client
• WebSocket — bidirectional streaming + in-band commands

Runner Protocol (plane ↔ runner)

Private gRPC contract for workers. Long-poll for work, stream events, report status, heartbeat, watch cancels. Small RPC surface on purpose; resources during a run (connectors, store, vectors) go over short-lived HTTP to the plane’s internal routes, run-bound.

• GetJob — long-poll assignment
• StreamEvents — progress back to the plane
• ReportStatus — terminal outcome
• Heartbeat — liveness + superseded signal
• WatchCancels — cancel stream

Why not one protocol?

Clients need a portable, SDK-friendly runs API. Runners need leases, fencing, and connector minting. Collapsing them either leaks internals to clients or starves workers of the lifecycle primitives a control plane must own. See why Agent Protocol on the front.

Also: opaque checkpoints (ThreadState ≠ LangGraph checkpointer) · generation fencing