Agentic Engineering Patterns: New Chapter on Git Workflows for Coding Agents
Simon Willison has been building out one of the most practical public references on agentic engineering, and the latest addition to his guide covers something that most coding agent workflows treat as an afterthought: Git. The new chapter makes a compelling case that version control isn't just a hygiene step at the end of an agent session — it's the highest-leverage primitive for reversibility, context loading, and safe parallel experimentation that most developers are leaving on the table.
The techniques range from foundational to clever. Using branch-per-task as an agent isolation primitive means a runaway session can always be undone with a single command. Seeding new sessions by pointing the agent at recent git log output lets it load meaningful project context without reading entire files — a significant token-efficiency win. More unusually, the chapter covers using git bisect as an autonomous debugging instrument: the agent writes the test, runs bisect, and identifies the breaking commit without human intervention. Commit granularity matters too — frequent small commits let you inspect agent progress at any checkpoint, not just at the end.
This chapter is part of a broader guide that has accumulated over 1,100 Hacker News points and 900 comments — a signal that it's resonating well beyond the usual AI-adjacent audience. For anyone running coding agents on real codebases, the Git chapter is a concise, immediately actionable read that reframes version control from a passive record-keeping tool into an active component of the agent workflow itself.