Orchestration
Agents overview · Orchestration overview · ① Session custody →
Implementation guides for Plane ② of the Agents Blueprint. After Plane ① freezes the route and async-starts the run, the agentic app copies that freeze into the run pin and executes the loop. Policy enforcement lives in PGAR Runtime.
THE CLAIM
The agentic app orchestrates; the LLM proposes steps; the PEP permits side effects. This is not intent routing and not model routing.
Four playbooks
| # | Playbook | One-line purpose |
|---|---|---|
| ① | Session custody | Route copy, run pin, durability store (not always Temporal) |
| ② | Memory | Isolation, TTL, retrieve_only, writes, erase |
| ③ | Autonomy shape | Run Pattern 0-3 from the route: single call, loop, fixed stages, or hybrid |
| ④ | Inference handoff | Every plan or synthesize call goes through Plane ③, then proposal gates at PEP |
Recommended path
- Wire agentic app: Plane ① handoff, front door, event bind, clarify/abstain
- Session custody: copy the route freeze; run pin; restart by
run_id - Memory: isolation, TTL, recall pack
- Autonomy shape: which pattern the route allows
- Inference handoff: gateway call, proposal check, PEP
- PGAR Runtime: SARAC, PEP/PDP, audit
Bridge reading: Agents Blueprint · PGAR Agentic app · Eval Action plane · Eval Tool plane.
What this plane decides
| Decides | Does not decide |
|---|---|
| Next tool proposal inside a scoped manifest | Which workflow or manifest is active (Plane ①) |
| When to call PEP, validation, step-up | Which LLM endpoint serves the call (Plane ③) |
| Session state, token custody, loop bounds | Entitlements at ingress (IdP + Plane ①) |
Who should read what
| Role | Start with | Then |
|---|---|---|
| AI platform | Session custody, Memory, Inference handoff | PGAR Agentic app, PEP enforcement |
| Domain squad | Autonomy shape | Agents Blueprint, Manifest registry |
| Governance | Inference handoff | Eval Action plane, Audit and replay |