A living, curated reading list for engineers shipping AI to production. Organized by pillar. New entries added as we read them.
Fresh issues from our own Substack — auto-pulled from the feed. See all issues →
OpenClaw isn’t a product to adopt. It’s a reference architecture to decompose. Five primitives, three production-grade use cases that earn real revenue, and a harness audit checklist for anyone build
Enterprise software only creates value when it’s actually deployed, and deployment is overwhelmingly a labor problem, not a software problem.
Vishnu Purohitham’s four shipped projects are a worked example of Model Reliability Engineering — and a soft hit on most of the AIfolio.
The launch numbers were one story. The production patterns are a different one.
A teardown of the two-harness architecture — async sandboxes for writes, virtual filesystems for reads — and what it teaches about wrapping a model in production.
Field notes on the harness loop at Harvey, Hippocratic, Anterior, and Azure SRE — where production failures compound into skill without retraining the model.
Anthropic's reference taxonomy of agent patterns — when to build a workflow vs. an agent, what each pattern is good for, and the failure modes of premature autonomy. The starting point for any production agent conversation.
Seven production patterns — evals, RAG, fine-tuning, caching, guardrails, defensive UX, collect feedback — with cites and rationale. The reference Eugene's been updating since 2023; the canonical "what does ship-to-prod look like" map.
How to actually improve an AI product after the demo: build evals from real failures, cluster errors, iterate the prompts and the data. The closest thing the field has to a playbook for "we shipped, it kind of works, now what."
How to actually improve an AI product after the demo: build evals from real failures, cluster errors, iterate the prompts and the data. The closest thing the field has to a playbook for "we shipped, it kind of works, now what."
Anthropic's reference taxonomy of agent patterns — when to build a workflow vs. an agent, what each pattern is good for, and the failure modes of premature autonomy. The starting point for any production agent conversation.
Seven production patterns — evals, RAG, fine-tuning, caching, guardrails, defensive UX, collect feedback — with cites and rationale. The reference Eugene's been updating since 2023; the canonical "what does ship-to-prod look like" map.
Practical guide to designing the context that agents see — how to build inputs that survive long-horizon tool use without devolving into prompt-stuffing. Closes the gap between "this works in the playground" and "this works in production with real users."
The canonical survey of agent design — planning, memory, tool use, reflection. Older than most "agent" posts you'll see but still the reference frame everyone draws from. Read this if you've ever wondered why your agent loops forever.
MCP — the open protocol Anthropic shipped for connecting agents to tools and data sources. Read this before designing any agent that needs to act on systems it doesn't own. Relevant tonight: how agents authenticate and enumerate capabilities at endpoints they've never seen.
Chip Huyen's early but durable map of the production LLM stack — what changes when you move from a notebook to real users. Cost, latency, hallucination, evals, drift. The starting reading for anyone moving from research to product.
Simon's yearly LLM survey — the most read 30 minutes you can spend before designing anything new. Tools matured. Agents didn't (yet). Per-token costs cratered. Required reading for grounding intuition.
Every few weeks, we ship a Reading Roundup issue — the best new items added to the library, with editorial notes. No vendor pitches, no idea-stage tourists.
Production-grade reads, in your inbox.
Free. No spam. Unsubscribe in one click. See past issues →