Handbook Agents
Overview
Handbook agents are package-owned under packages/domain-handbook/src/agents/** and use the shared agent factory pattern from @maf/platform-runtime/agents/shared/index.
The Handbook package currently registers seven agents:
| Agent | Profile ID | Registry ID | Functional group | Tools |
|---|---|---|---|---|
| Handbook Product Agent | handbook-product-agent | handbook_product-agent | Product Workflows | handbook-search-policy-evidence, handbook-generate-verified-answer |
| Handbook Source Context Synthesis Agent | handbook-source-context-synthesis-agent | handbook_source-context-synthesis-agent | Source Context | none |
| Handbook Policy Evidence Gate Agent | handbook-policy-evidence-gate-agent | handbook_policy-evidence-gate-agent | Policy Review | none |
| Handbook Policy Reconciliation Agent | handbook-policy-reconciliation-agent | handbook_policy-reconciliation-agent | Policy Review | none |
| Handbook Checklist Verifier Agent | handbook-policy-checklist-verifier-agent | handbook_policy-checklist-verifier-agent | Policy Review | none |
| Handbook Remediation Drafter Agent | handbook-policy-remediation-drafter-agent | handbook_policy-remediation-drafter-agent | Policy Review | none |
| Handbook Evidence Synthesis Agent | handbook-policy-evidence-synthesis-agent | handbook_policy-evidence-synthesis-agent | Policy Review | none |
Source: packages/domain-handbook/src/agents/index.ts.
Handbook Product Agent
The product agent is a thin adapter over Handbook product-owned tools. Its prompt tells the agent to use only attached Handbook tools for product data and not to invent policy evidence.
Primary use cases:
- inspect and run product workflows
- search policy evidence through the product retrieval contract
- generate verified answer packages grounded in retrieved evidence
- preserve product contract boundaries when called from Admin or route adapters
Source: packages/domain-handbook/src/agents/handbook-product-agent.ts.
Source Context Synthesis Agent
The source-context agent converts cited handbook evidence into structured,
provenance-preserving data-point answers. The
handbook.extract-source-context.v1 workflow persists those answers as source
context memory so later generation and checklist review can reuse uploaded
handbook facts without treating them as authoritative policy.
Source: packages/domain-handbook/src/agents/handbook-source-context-agent.ts.
Policy Evidence Gate Agent
The evidence gate agent is scoped to policy review. It decides whether retrieved MSBA policy evidence is plausibly relevant to a handbook segment.
It should:
- select policy evidence only when it can govern, contradict, permit, or constrain the handbook language
- reject weak or off-topic evidence
- treat missing evidence as unresolved review work, not proof that the handbook is wrong
- return only the requested structured JSON
Source: packages/domain-handbook/src/agents/handbook-policy-review-agents.ts.
Checklist, Remediation, And Evidence Synthesis Agents
The remaining policy-review agents separate three reviewer-facing concerns:
- the checklist verifier compares one checklist requirement with cited handbook evidence and reports satisfied, missing, or unverifiable coverage;
- the remediation drafter proposes human-reviewable handbook wording for missing or partial findings and must not present draft text as evidence;
- the evidence synthesis agent turns already-selected evidence into clear reviewer blocks while preserving caveats about weak, truncated, or heading-only evidence.
All three are tool-less structured-output agents. They receive evidence and contracts from the policy-review runner rather than searching independently.
Source: packages/domain-handbook/src/agents/handbook-policy-review-agents.ts.
Policy Reconciliation Agent
The reconciliation agent compares cited handbook evidence with cited MSBA evidence after the evidence gate has selected plausible sources.
It should:
- mark
contradicts_policyonly for direct semantic conflict - avoid treating local detail as a contradiction unless MSBA expressly forbids it
- use
overstates_policyorneeds_manual_reviewwhen the evidence is ambiguous - ground suggested wording in cited evidence only
- return only the requested structured JSON
Source: packages/domain-handbook/src/agents/handbook-policy-review-agents.ts.
Registration And Discovery
buildHandbookAgents() configures and returns all seven agents in the Handbook agent registry. When the Handbook domain is enabled, runtime discovery exposes them through:
GET /agents.jsonGET /handbook/routes.json- generated
/handbook/agents/{agentSlug}/generateand/handbook/agents/{agentSlug}/streamroutes
Use /handbook/routes.json in the target environment for exact route slugs.
Related Docs
Route families that expose Handbook agents, tools, workflows, and Admin BFF behavior
Where evidence-gate and reconciliation agents fit in the review pipeline
Shared profile and prompt override mechanics
Shared descriptor/factory pattern used by Handbook agents