Book roadmap¶
The ten chapters follow a single engineering chain. Each chapter addresses a question that cannot be skipped and turns its conclusions into reviewable contracts, code, and runtime evidence.
Part I: architectural judgment and execution control¶
First decide whether the system needs autonomy, then place model-proposed actions in a controlled runtime, and finally address control and state ownership across multiple agents.
The essence of agents and architectural boundaries
Distinguish ordinary code, LLM workflows, single-agent systems, and multi-agent systems to determine where dynamic routing and execution autonomy should stop.
Deliverables: agent decision framework, autonomy boundary matrix, ADR 02Tool calling: state machine, control plane, and MCP
Turn a Tool Call from a model suggestion into execution that is authorized, recoverable, auditable, and does not repeatedly produce side effects.
Deliverables: Tool Contract, state machine, Checkpoint, MCP boundary 03Multi-agent collaboration: architectural patterns, protocols, and incident response
Choose among Handoff, Supervisor, Router, Fan-out, and Join based on control, responsibility, state, and failure semantics.
Deliverables: pattern decision record, delegation contract, deterministic JoinPart II: trustworthy context and production assurance¶
Ensure the system uses only information qualified to serve as evidence and can still uphold business boundaries when infrastructure fails or the model is misled.
Context engineering: Agentic RAG, GraphRAG, and evidence chains
Govern queries, hybrid retrieval, relational evidence, permissions, temporality, and citations so that Context becomes a traceable runtime product.
Deliverables: Context Pack, retrieval gates, GraphRAG evidence paths 05Production foundation: deployment, observability, and failure recovery
Clarify state ownership, health semantics, telemetry correlation, reliability safeguards, and recovery order to turn “it starts” into runtime guarantees.
Deliverables: production-readiness contract, SLO, recovery drills, and evidence 06Agent security: defense in depth, Tool Guard, and red-team validation
Assume the model may fail, and limit the real-world impact through identity, the intersection of permissions, resource binding, data release, and red-team regression testing.
Deliverables: Tool Guard, security audit, privacy lifecycle, and red-team evidencePart III: system integration and continuous delivery¶
Combine distributed components into a single causal chain, use evaluation and production operations for continuous validation, and ultimately deliver a reproducible, traceable, and maintainable system.
System integration: Supervisor, A2A, MCP, and Context Graph
Connect goals, plans, delegation, tools, state, evidence, errors, and recovery semantics to create an end-to-end system with verifiable acceptance.
Deliverables: layered architecture, Context Graph, system integration contract 08Quality engineering: Golden Dataset, layered evaluation, and continuous regression testing
Evaluate the complete causal chain from Goal to Answer, using deterministic checks, Judge, N-run, and gates to control stochasticity.
Deliverables: Golden Dataset, layered metrics, continuous regression gate 09AgentOps: incident diagnosis, system resilience, and cost governance
Reconstruct the causal chain of an incident around the user's goal, and execute audited controls, recovery, cost attribution, and continuous optimization.
Deliverables: goal-level SLO, Incident Bundle, GameDay, and ORR 10System delivery: acceptance evidence, Release Gate, and open-source governance
Move from requirements traceability to reproduction in a clean environment, then complete the public delivery with an SBOM, build provenance, and governance files.
Deliverables: Evidence Package, Release Gate, verifiable v1.0Project throughout the book: CaseOps¶
The entire book uses the same complex claim-handling case: investigate why case C-102 was not completed and confirm the missing materials and applicable rules; if the customer needs to be contacted, the system may only generate a draft and must not send it automatically.
CaseOps begins with a deterministic domain core and progressively gains a tool state machine, multi-agent collaboration, governed Context, production runtime guarantees, security controls, system evaluation, AgentOps, and verifiable releases. The complete implementation is maintained in the separate CaseOps code repository, with v1.0.0 as the official release.