Overview

AI coding agents are fast. Durable engineering systems are deliberate for a reason.

This playbook helps teams treat agentic programmers like durable workers: isolated by branch, checkpointed by atomic commit, verified before handoff, and stopped by risk gates before they can damage production trust.

The aim is not to make agents busier. The aim is to make their work survive review, rollback, incident pressure, and maintainer scrutiny.

What You Get

global agent operating policy
repo-level AGENTS.md template
PR template
review pack template
blocker template
risk escalation checklist
verification checklist
examples for common repo types

Durable Agent Concepts

branch = isolated execution context
atomic commit = checkpoint
review pack = execution log
verification = health check
rollback plan = recovery path
risk escalation = circuit breaker
dry-run and idempotency = safe retry semantics
pull request = human review boundary

The Default Loop

state the objective
state the expected blast radius
create a branch from latest main
make the smallest coherent change
review git status and git diff
stage only intended files
run relevant verification
commit one reviewable intent
rebase before PR
open a focused PR
return a review pack

The Standard

The standard is not “agents make lots of commits.”

The standard is durable engineering throughput: small branches, reviewable checkpoints, clear verification, explicit risk, and no production surprises.