Agent Engineering HandbookGitHub

Follow the guides back to the source

Three reports distinguish inspected code, author-reported results, remote CI records and local runs. Each separates open pull requests from merged work and links pinned source snapshots where possible. Work under an account is not necessarily written by that person.

REPORT 01 · 10 GUIDES CAME FROM THIS · 2,457 WORDS

Theo: T3 Code and Melee

Ten of the thirteen guides draw on the video and this investigation of T3 Code and Melee. The report connects advice to inspected code.

  1. A completion contract specifies the real output, how complete its inputs must be and the outcome its consumer sees.
  2. Recurring mistakes need checks at existing lint, type, test and build boundaries that reject bad changes and accept valid ones.
  3. Explicit ownership separates agent writes and test state, contains provider quirks in adapters and checks remote access to artifacts.
  4. Project decisions stay beside relevant code, while evidence distinguishes local runs, remote CI, author reports and untested hardware.

Read the full reportEditable Markdown

REPORT 02 · 2 GUIDES CAME FROM THIS · 1,646 WORDS

Matt Pocock: Course Video Manager

A glossary and one HTTP transport give this codebase a shared language. Its source informed two guides on domain language and durable artifacts.

  1. Shared vocabulary, lasting decisions, bounded modules and callable agent APIs inform the guide to speaking a codebase's language.
  2. Output identity, resumable work and recovery from malformed results inform the guide to artifacts that survive interruptions.
  3. The report recommends making the intended outcome easy to request and hard to misreport.

Read the full reportEditable Markdown

REPORT 03 · 1 GUIDE CAME FROM THIS · 1,122 WORDS

Boris Cherny: public engineering work

Guide 13 adapts five validation methods: fixtures, built-artifact smoke tests, fuzzing, real-world schemas and conformance checks. Tests were inspected, not run locally.

  1. Guide 13 adapts layered validation to test public output, reproduce new failures and preserve meaningful baselines.
  2. Before-and-after measurements compare the same work to establish whether a change improved the result.
  3. Existing lessons on lint, adapters and native loading provide supporting references for the four skills.

Read the full reportEditable Markdown