Skip to content

What's New

This page is a short reader-facing log of major additions to the book and reference runtime. It is not a replacement for git history; it exists to show that the project is active and what is already here.

Current as of April 29, 2026.

Book

Part VIII on the lifecycle of agent systems

The book now includes a full block on SDLC -> ADLC, change management, assurance loops, supply chain, retirement, misalignment, behavioral evals, AI-native observability, and inventory control.

Why it matters: the site now covers not only architecture and rollout, but the lifecycle of an agent system after release.

A stronger production contour across Parts I-V

The book now includes sharper bridges between architecture, retrieval, execution, and eval discipline:

  • Part I now separates runtime architecture more explicitly from the training layer and the product surface;
  • Part II now gives a clearer taxonomy for prompt injection, jailbreaking, and action hallucination;
  • Part III now strengthens the retrieval contour with semantic gap, HyDE, RAG first, and a clearer distinction between continued pretraining and SFT;
  • Part IV now adds practical guidance for large tool catalogs, semantic tool filtering, and explicit MCP host / client / server roles;
  • Part V now adds a stronger product framing for latency budget and a more practical treatment of LLM-as-a-judge.

Why it matters: the book now covers not only the baseline platform layers, but also the everyday production questions that tend to surface between design review, eval loops, and rollout.

Reference

A reusable reference layer

The site now includes reference pages for:

  • traces and event catalog;
  • eval datasets and grading contracts;
  • policy bundles and approvals;
  • change review and rollout gates;
  • lifecycle artifacts;
  • memory retrieval contracts.

Why it matters: readers can now move directly from explanatory chapters to reviewable schemas and contract artifacts.

Runtime

Runnable reference runtime

The repository includes agent_runtime_ref, a small runnable package that now supports:

  • approvals and delegated authorization context;
  • controls and lifecycle runtime-control inspection;
  • lifecycle artifacts;
  • session export and replay summaries;
  • eval dataset export;
  • trace export with redaction, redacted summaries, replay preservation, and schema versioning.

Why it matters: the book is now backed by runnable code, not only narrative documentation.

Practical Appendix

A stronger practical appendix

The site now includes:

  • a glossary;
  • cheat sheets;
  • case studies;
  • policy templates;
  • a research frontier page;
  • a community roadmap.

Why it matters: readers have fast access to checklists, case studies, glossary entries, and practical assets without reading the entire book linearly.

Stronger entry pages

Updated:

Those pages now make it easier to find short paths into topics such as:

  • semantic tool filtering;
  • HyDE and RAG vs training;
  • latency budget and routed pipelines;
  • LLM-as-a-judge and judge calibration;
  • the difference between prompt injection, jailbreaking, and action hallucination.

Why it matters: the new topics are now visible not only inside individual chapters, but also at the reader entry-point level.

Publish Readiness

A cleaner site before publication

Before the next deploy, a publisher-facing quality pass was started and moved forward:

  • draft and planning pages are excluded from the published site and sitemap;
  • OpenGraph/Twitter metadata and a social preview image were added;
  • the search index, sitemap, robots file, local assets, anchors, alt text, and external links were checked;
  • the READMEs in all three languages now include a fast-forward publish checklist for main and docs-prod.

This does not mean the publisher-facing layer is fully closed: deep EN/ZH QA, rendering/export QA, sample-chapter polish, and manuscript/online-companion packaging still remain.

Why it matters: the published site should keep moving toward a polished reader-facing product, not feel like a raw build of Markdown files.

What this means for readers

  • You can use the book as a handbook.
  • You can reuse the reference pages as engineering starting points.
  • You can run the example runtime instead of reading only Markdown.
  • You can anchor the architecture in recent material from OpenAI, Anthropic, Google, Microsoft, and NIST.

Continue