Practitioner's guide to AI coding agents, distilled from 75+ research references, 6 real PR transcripts with 43 documented failures, and a full overhaul validated against live codegen benchmarks.
Delete rules, enforce mechanically, verify independently, stop when converged.
How rules, skills, tests, benchmarks, and feedback loops combine to let an AI agent implement software tickets end-to-end — and get smarter with every iteration.
This document explains how this repository is set up to enable reliable, automated “agentic” workflows—where an AI assistant can take a high-level request like “complete CAN-242” and drive it to a verified pull request with guardrails.
It is written for non-technical readers first, with technical detail in later sections.
This document explains how this repo is set up to enable an automated, repeatable agentic software workflow (from “work starts” → “PR is created” → “quality gates pass” → “release is automated” → “learnings are captured so the next run is better”).
It’s written for non-technical readers first (plain language, metaphors), while still being grounded in the actual repo artifacts (rules, skills, docs, tests, CI, benchmark harness).