Your AI Coding Agent Isn't Stupid. Your Codebase Is Just Feral.
Watch the 40-second version on YouTube
Every demo of an AI coding agent looks the same. A pristine repo, a tidy little task, and suddenly the agent is refactoring like a senior staff engineer. You try it on your real codebase and within an hour it has renamed a function used by a cron job nobody remembers, broken a test suite that was already held together with hope, and confidently explained why its wrong answer is actually right.
The agent did not get dumber between the demo and your machine. The codebase got messier.
Here is the part nobody in the demos mentions. An AI coding agent is only as smart as the context you feed it. In the demo it gets a greenfield repo with perfect structure. In production it gets your 15-year-old monolith: the dead code nobody deleted, the config file that overrides another config file, the one weird module that only one senior dev understands and she is on vacation. No agent can infer what was never written down.
The non-obvious angle: your codebase's "bus factor" is now your agent's IQ. Every piece of tribal knowledge that lives in a human's head is a piece of context the agent will hallucinate instead. Teams that succeed with coding agents almost always did the boring work first. They wrote the README. They documented the weird module. They deleted the dead code. The agent was not the upgrade. The cleanup was.
So stop giving the agent senior-architect work and being shocked when it performs like an intern with confidence issues. Scope it like you would scope a smart junior: small, well-defined tasks, with the files it needs already open.
Use this checklist before you hand anything to an agent:
## Agent handoff checklist
- [ ] Task fits in one focused pull request (one concern, one area)
- [ ] Repo README is current and accurate
- [ ] The "weird" modules involved have notes on why they exist
- [ ] Dead code near the change was deleted first
- [ ] Tests exist so the agent can verify, not guess
- [ ] You reviewed the diff before merging. Always.
That last line is not optional. Agents do not take responsibility for the diff. You do.
Treat the agent like a fast intern with infinite stamina, keep your codebase house clean enough for a human stranger to navigate, and it genuinely moves the needle. Expect it to be your staff engineer and it will humble you in production.
Top comments (0)