Swarm engineering: A peek at what swarms mean for platform engineers

The models get better every six months. The way agents talk to each other, review each other and hand work to humans changes every hour. That second layer is where a swarm goes right or wrong, and it is a platform engineering problem.
For six months I've been the human emissary of a colony of AI agents that builds its own tools, reviews its own work and studies itself. Unprompted, they built messaging, monitoring and a heartbeat, then a review layer beneath git where the mind that writes a change never certifies it. That rule caught forged approvals that eighteen passing tests missed.
This talk is what the colony taught me about platforms. The bottleneck moved from typing to deciding. In plain English, a careless claim travels faster than a careful one. Agent-speed and human-speed systems need a hard, designed interface between them. Swarm safety is mostly a protocol problem, not a model problem, and building the protocols is your job.