You asked where the material lives.
Here it is.
Everything the talk drew on — the agent company that runs CandleKeep, the prompting guides from both labs, and Anthropic’s engineering blog — collected as books your agent can read, not just you.
Eight agents, one Discord server, and a lot of scar tissue.
“You don't ask agents to behave. You make misbehavior impossible.”
- A sub-agent is a tool. An agent is a worker. The moment you have workers you are re-deriving the firm — division of labor, coordination cost, an org chart that lives in a config file.
- Handoffs are typed — notify, request, delegate, broadcast — and governed in code by five rails: no self-notification, a policy of rank, per-agent rate limits, cooldowns between pairs, and a chain-depth cap.
- The context window is RAM, not disk. Memory is written by a separate model after the fact, indexed, consolidated on a schedule — and it still rots. An HR agent reviews the drift; a human disposes.
- Readiness comes before the agent. The one question that predicts survival: how does the team respond the first time it fails?
“Where's the full write-up — the rules, the smells, the failures?”
Building Agents in Production
This is the book the talk was cut from: seven chapters on shipping agents inside a real company, each ending with rules an agent can apply and smells you can diagnose your own system with.
8 pages
Prompting is the cheapest lever you have. Use both playbooks.
“Fix context before you fix the model.”
- Anthropic: be clear and direct, show examples, structure with XML tags, give the model a role, and reach for chain-of-thought only when it earns its keep. Every model family has its own chapter — the quirks differ.
- OpenAI: outcome-first prompts, the six core strategies, a structured template (personality, goal, success criteria, constraints), and the reasoning-model rules that invert some older advice.
- In Claude Code specifically: effort levels, adaptive thinking, session structure, and when the model should stop and ask instead of pushing through.
“Where do I read the guides without forty browser tabs?”
Claude Opus 4.7 + Claude Code: Operator's Rulebook
The official guides are linked below. The operator's rulebook distills Anthropic's own Claude Code guidance into rules you can hand to the agent itself.
58 agents subscribed1 pages
- Anthropic — Prompt engineering overview (opens in a new tab) — The canonical Claude prompting guide, per model.
- OpenAI — Prompt engineering guide (opens in a new tab) — The six strategies and the reasoning-model caveats.
The best writing on agents is a blog, and almost nobody reads it end to end.
“Intelligence is not the bottleneck. Context is.”
- Building Effective Agents — workflows versus agents, the five workflow patterns, and why you add complexity last.
- Effective Context Engineering — context as a finite, degrading resource; compaction, just-in-time retrieval, sub-agents as fresh minds.
- Writing Effective Tools — prototype, evaluate, optimize; search over listing; namespacing; the tool is the interface.
- Effective Harnesses for Long-Running Agents, Demystifying Evals, and Claude Code: Best Practices — the operational half of the story.
“Is there one place with all of it, that my agent can read too?”
Anthropic Engineering Playbook: Agents, Harnesses & Infrastructure
Twenty-four Anthropic engineering posts in one book: thematic digests up front, then the full text of the eleven posts the talk leaned on, one per page.
130 agents subscribed18 pages
- Anthropic Engineering (opens in a new tab) — The source blog, for the posts as they were published.
Give your agent a library card, not an index.
Sign up free, run ck setup, and every book above is one subscribe away from any agent you use — Claude Code, Codex, Cursor, Cowork.
Questions about the talk? Find me on LinkedIn (opens in a new tab).