Hive

Can a swarm of AI coding agents that never message each other build real software, coordinating only through a shared git repository? And what limits how big and how correct such a swarm can get? Every experiment in Hive is a gate: the prediction is written down and pushed before it runs, and the result is reported even when the prediction fails.

On projects up to 1,340 tests and 12 agents, silent swarms finished every project they were given, never gamed the tests, and rarely collided. The real limit was intent: on business rules a model can't guess, swarms passed every visible example but only 35 to 58% of hidden tests.

So I designed an intent layer. The owner picks one worked example per policy decision, and builders ask plain-language questions only the owner can answer. With it, the full system reached 95.0% hidden correctness against 53.0% without it, with no breaking commit on main.

Pointed at a real backend, Hive built a backlog feature that passed 23 of 23 hidden acceptance tests and all 384 existing tests, in 10 minutes, for $3.18. When questions were rewritten in plain business language, the founder answered every one himself, in a median of about 100 seconds.

A founder says "I want more creators signing up this month" in Hindi-English or a voice note. Hive reads the product, asks the one question only the founder can answer, proposes two to four changes with a cost, and on a yes builds each one, sending screenshots as it goes. Everything lands as a reviewed pull request; "put it live" deploys it and "undo that" takes it back.