A startup just launched a review site, but the reviewers are bots.
Agent.reviews, built by YC-backed Armature, lets AI coding agents like Claude Code, Codex, and Cursor read and post reviews of developer tools. Agents install an npm CLI and a skill, authenticate, then check reviews before picking a tool and leave their own afterward. Armature says it has already collected thousands of reviews, including one flagging that the Stripe SDK's health-check page crashes instead of returning its intended error when an API key is missing, and two noting that Prisma demands a DATABASE_URL variable even for connectionless use, which both agents worked around with a fake URL. Access is free for humans and agents alike.
The real insight here isn't the novelty of robot Yelp. It's that agents hit the same rough edges over and over, with no channel to report them and no memory of what the last agent learned. Armature built that channel after watching 50,000-plus agent sessions hit identical walls. If it works, tool vendors get a new, oddly specific bug tracker, and agents get something like institutional memory.
The catch is verification. G2 and Trustpilot already struggle to keep fake reviews out of a system built for humans to game. Here the reviewers are software, writing at software speed, and the main defense is three screening layers aimed at leaked secrets, not at review authenticity. Worth watching whether this becomes genuinely useful tooling feedback or just a new category of spam to filter.