85% of Engineers Are Running Dozens to Hundreds of Agents; the Comments Ask Who Reviews Them First

A tweet on X highlighted a share from the Claude Code lead. In the video, the lead gives a number: 85% of engineers are running dozens or hundreds of agents. The tweet text is enthusiastic, saying they have taken over the AI space and telling people to watch the video and learn. The original post is here.

The number is striking, but the comments are closer to the engineering reality. Biscuit asks who reviews the output of hundreds of agents. Nash says running dozens of agents is the easy-to-replicate part; after a few, all your time goes into reviewing what they did. Nour Salama is already doing a small-scale version: one git worktree per task to keep agents from colliding with each other. He follows up by asking how all those branches eventually get merged back.

These replies expose the crux of multi-agent workflows. The marginal cost of generating code is falling, but review, merging, conflict handling, and accountability are not falling in tandem. One worktree per task can isolate file changes and reduce low-level collisions, but semantic conflicts, duplicate implementations, missed tests, and the final merge strategy still require human decisions. The more agents there are, the longer the review queue gets, and the more merging looks like the bottleneck.

For AI practitioners with a development background, the parts worth copying are task decomposition, workspace isolation, review queues, and merge rules. For ordinary users, when you see dozens or hundreds of agents, do not mistake it for autocomplete at first. It is more like spreading work out in parallel and then having humans clean up the results. The questions in the comments are closer to everyday reality than the excitement of the tweet.