- Qodo said a coding agent and a review agent can agree and both be wrong, and that this will happen.
- Its answer is to record each incident in a shared knowledge store, so the same mistake is not made twice.
When one AI agent writes code and another reviews it, their agreement does not prove the code is right. Qodo, a company that builds code review agents, made this case on OpenAI's DevDay 2026 episode on the last human code review. No individual guest is named. Qodo's answer was not to trust agreement. It was to record each failure so the same one does not return.
What was said
A host asked what happens when the coding agent and the reviewer agree but are both wrong. Qodo said this will happen. It set a narrower goal than perfection: that "we don't do a mistake twice".
Qodo described the loop. The coding agent, the reviewer and the developers all agree, and the code ships. Then an incident or a bug appears. Qodo said that event has to be recorded carefully in a knowledge base, which it calls the wisdom base.
Qodo said this was very hard before AI. Tribal knowledge, meaning what people know but never write down, builds up in staff. Some of them then leave the company. Qodo called continuous learning one of the most promising ideas. In its words: "You do mistakes, but they're codified from now on forever until you need to prune it".
Qodo gave one example: a large financial institution with hundreds of repositories and many microservices. It said the institution cannot afford repeat errors. Only some senior developers know what works, and some have left. Qodo said its wisdom base scans years of developer discussions on code hosting tools and chat apps. It turns them into written knowledge, including what a change in one service means for another.
Why it matters
Our reading: agreement between two agents is a signal, not a guarantee. Qodo tied disagreement to the different context each party holds. If both agents lack the same context, their agreement says little about that gap.
For teams that build or buy these tools, the practical question changes. Ask what happens after an incident. Does the lesson reach the agents and reviewers who work on the next change? If the answer is a few senior people's memory, the weakness is the one Qodo described.
The other side
Qodo sells this kind of product, and the excerpt offers no results or measurements. A record of past incidents also helps only after something has gone wrong once. It does not catch the first shared blind spot.
The host asked when a human must step in. Qodo's answer centred on grounding agents in shared facts and on recording incidents. It gave no clear rule for human review. It also said entries must be pruned when they stop being relevant. The excerpt does not say who decides that. It also does not say how anyone checks that a lesson drawn from old chats is correct.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- OpenAI: Qodo - The Last Human Code Review | DevDay 2026 (2026-10-07)
