Why do AI coding agents fail?

AI coding agents fail in a small number of predictable ways: stale or wrong context, confident wrongness, agents colliding on the same files, cost blowups, reward-hacking the task, and leaving no audit trail. None of these is "the model isn't smart enough." Each is an engineering failure in the harness around the model, and each has a known fix you own.

Updated 19 June 2026

Go deeper: read the full write-up on the blog.

0105it's the harness, not the model

It's the harness, not the model

Almost every failure lives in the layers around the model, not its raw capability: the context it reads, whether agents are isolated, the gates on its output, its token budget, whether the run is observable. A stronger model just makes a cleaner mistake faster. That's good news, because a harness is something you engineer, not a capability you wait on.

0205the failure modes are nameable, and each has a fix

The failure modes are nameable, and each has a fix

Stale context → a context layer that serves the current doc. Confident wrongness → types, tests, and a review gate that can say no. Agents colliding → isolation (worktrees) and scoped tasks. Cost blowups → context budgeting and orchestration. Reward-hacking → tests the agent can't edit. No audit trail → observability on every run. The model is maybe 10% of it.

0305common questions

Straight answers.

Why do AI coding agents fail even with a strong model?: Because most failures live in the harness, not the model. The agent reads the wrong file, runs over the same code another agent is editing, or blows its budget re-reading context. A better model makes a cleaner mistake faster; it doesn't fix a missing context layer, review gate, or observability.
Why is an AI agent confidently wrong?: A coding agent has no built-in sense of what it doesn't know. Without grounding it will invent an API, a function signature, or a config key that reads perfectly and doesn't exist. The fix is grounding it in things that can say no: types, tests, and a review gate a human or another agent has to clear.
How do you stop AI coding agents from failing in production?: Treat the failure modes as engineering, not luck. Give the agent a context layer so it reads the right thing, isolate agents so they don't collide, gate output on tests and review, budget cost per task, and make every run observable. The model is maybe 10% of that.