Reference Layer¶
If the book explains why a safe agent system should be built this way, the reference layer helps answer what exactly should be captured in artifacts, schemas, and contract pages.
This layer is intentionally supportive, not primary. It exists to anchor the book's argument in reusable engineering materials, not to replace the reader journey of the book itself.
Use this section when you need to:
- find the right contract page quickly;
- prepare a design review or rollout review;
- extract reusable artifacts for your team;
- move from conceptual chapters to applied engineering materials.
If you are new to the project, start with the book first. Come here when you want the supporting schemas, checklists, and contract surfaces behind the main argument.
What this layer does not promise:
- it does not replace the book's reader journey;
- it does not explain the main causal argument chapter by chapter;
- it does not try to be the primary place where trade-offs and layer boundaries are learned.
Start here¶
For a short path in, start with:
Support-triage artifact route
If you read the book through the support-triage case, keep the trace, eval dataset, policy bundle, approval record, incident record, change rollout, lifecycle artifact, and registry operations pages next to it. Those contracts turn the duplicate-ticket incident from a story into a reviewable artifact set.
Canonical case artifacts
The three canonical cases enter the reference layer through different artifact sets. Support triage depends on the approval record, policy bundle, trace schema, and duplicate-ticket recovery evidence. Internal knowledge assistant needs the memory/retrieval contract, freshness checks, access control, and knowledge provenance. Incident coordination connects the incident record, escalation evidence, notification side effects, response ownership, and post-incident learning.
Safe-agent schema spine
For a short route through safe-agent architecture, keep the trace schema, eval schema, and memory/retrieval schema together. They now connect the MCP threat model, A2A handoff trust contract, verifier verdict record, governance action record, memory poisoning review fields, and unified agent threat evidence.
Schemas and contract pages¶
- Trace Schema and Event Catalog
- Eval Dataset Schema and Grading Contract
- Policy Bundle Schema and Approval Contract
- Approval Request and Decision Record Schema
- Incident Record and Postmortem Linkage Schema
- Change Review and Rollout Gate Schema
- Lifecycle Artifact Schema
- Memory Record and Retrieval Contract Schema
- Causal Debugging and Root-Cause Analysis for Agent Systems
- Memory Eval Patterns for Agent Systems
- Tool Failure Recovery Patterns for Agent Systems
Practical pages¶
- Reference Package
- Case Studies
- Policy Templates and Checklists by Use Case
- Incident Response Playbook for Agent Systems
- Handbook for Agent Registry and Inventory Operations
- Postmortem Template for Agent Systems
Fast Topic Routes¶
If you do not need the whole reference layer, but only a short path into one concrete topic, start here:
- Tool catalog design, semantic tool filtering, and read/write taxonomy: Chapter 8. Execution Model and Tool Catalog
- MCP host/client/server roles, capability transport, and sandbox boundaries: Chapter 9. Sandbox Execution and MCP as an Integration Contract
- Semantic gap, HyDE, and RAG vs training: Chapter 7. Retrieval, Compaction, and Background Updates
- Latency budget, fast path / slow path, and routed pipelines: Chapter 12. SLO for Agent Systems
- LLM-as-a-judge, calibration, and judge-human agreement: Chapter 13. Offline Evals, Online Evals, and Regression Gates
Continue¶
The simplest rule is:
- use the book for argument and sequencing;
- use the reference layer for support artifacts and implementation-facing detail.