All work
AI experiments · Sep 2026
Fleet simulator
Drag from 100 to 300,000 concurrent AI agents and see what it takes: model calls, tokens, cost, sandbox servers, monitor decisions, and human reviewers. Companion to a 4,400-word architecture post.
agent-fleet.shrid5.workers.dev
01The model
Calibrated to Anthropic's public numbers: about 30,000 agents and a billion monitor decisions a month. Little's law gives an 80-second agent step, and the rest falls out: cost per agent-hour, rate limits, servers, and reviewers.
02The thesis
You can buy more GPUs, sandboxes, and tokens. Human review is the one resource you cannot buy more of, so the whole design bends around it.