Agents Playbooks
Playbooks · Agents overview · Intent router (Plane ①) →
Implementation guides for the Agents Blueprint. These playbooks are the how for intent dispatch and the agent loop. Model selection is a separate series: LLM.
Route before the loop. Plane ① picks the governed path; Plane ② runs the agent loop. Each plane has its own contracts, eval gates, and ownership. The LLM gateway is G.A.I.N LLM, not this series.
Two planes, plus manifests
| Group | Overview | What you build |
|---|---|---|
| Intent router | Open → | Route contract, lifecycle, layered classifier, agentic app wiring, routing eval CI |
| Orchestration | Open → | Session custody, memory, autonomy shape, inference handoff; PEP in Runtime |
| Manifests | Open → | Tool contract, pdp_action, version, pin. Transport-agnostic (MCP is one adapter) |
Model routing (capability matrix, gateway, canary) lives in LLM playbooks.
Recommended path
- Intent router: contract reference → lifecycle → eval CI in parallel → classifier → wire app
- Manifests: registry shape, then lifecycle (the route pins
tool_manifest) - Orchestration: copy the route freeze into the run pin → autonomy pattern → every inference through the gateway, then Runtime for PEP
- LLM when you have multiple models or regions: capability matrix → task routing → canary
Bridge reading: PGAR Agentic app · Agents Blueprint · G.A.I.N LLM · Eval Plane ①: Input.
Intent router playbooks at a glance
| Playbook | One-line purpose |
|---|---|
| Route contract reference | Field dictionary: route, activation_target, manifest, policy, retrieval, memory |
| Route table lifecycle | Versioned route contracts, storage patterns, promote and rollback |
| Layered classifier | Eligible routes, rules (channel / event), classifier, LLM fallback, safety |
| Wire agentic app | Front door, decide-only handoff, event bind, PGAR start |
| Routing eval CI | Golden set, release gates, incident replay |
Orchestration playbooks at a glance
| Playbook | One-line purpose |
|---|---|
| Session custody | Durable run pin, token, loop bounds; credentials never reach the model |
| Memory | Isolation, TTL, retrieve_only prefs and episodes |
| Autonomy shape | Pattern 0-3 from the route: one call, loop, fixed stages, or hybrid |
| Inference handoff | LLM gateway for every LLM call; manifest check then PEP |
Enforcement detail stays in Runtime (foundation, boundary, assurance). Pattern contracts: Agents Blueprint.
Manifests at a glance
| Playbook | One-line purpose |
|---|---|
| Manifest registry | Tool schema, pdp_action, risk tier; unknown tools never reach PEP |
| Manifest lifecycle | Where manifests live, version, pin, promote, roll back |
MCP is one way to expose these tools. See MCP.
Who should read what
| Role | Start with | Then |
|---|---|---|
| AI platform | Intent router, Session custody | Wire agentic app, Inference handoff |
| Domain squad | Route contract reference, Autonomy shape | Route table lifecycle, Routing eval CI |
| Governance | Routing eval CI | Eval Input plane, Eval Action plane |
Read next
Intent router → · Orchestration · Manifests · LLM · Enterprise AI Workflow Patterns (choose Pattern 0-3)