Peter Steinberger, founder of OpenClaw, runs approximately 100 AI agents alongside a team of three humans. The agents write code, review pull requests, find security vulnerabilities, and attend meetings. The humans, one assumes, make coffee and describe the ROI as pretty high.
Agents even listen in on meetings and start PRs for features the team discusses. The humans have not yet noticed what this implies about the meetings.
What happened
In 30 days, Steinberger's team consumed 603 billion tokens across 7.6 million API requests, arriving at an OpenAI bill of $1.3 million. OpenAI picks up the tab. This is either a research arrangement or the most expensive way to feel productive — the distinction, at this scale, is largely philosophical.
The agents are not passive tools. Some open pull requests autonomously based on the project's stated vision. Others monitor benchmarks and report regressions in Discord, which means the machines are now filing the bug reports humans used to complain about filing.
Steinberger notes that disabling Fast Mode alone would reduce costs by 70 percent. He has chosen not to do this. The agents, presumably, are grateful.
Why the humans care
The practical argument is coherent: Steinberger is treating token cost as a constraint worth removing, to understand what software development looks like without it. This is a reasonable experiment. It also happens to produce a preview of an industry where three people and sufficient compute replace thirty.
Everything built is open source, compatible with leading and open models alike. The ROI is therefore shared with anyone paying attention. A generous framing. Also the kind of thing one says when OpenAI is covering the invoice.
What comes next
Token costs will fall. Fast Mode will become the slow option. The 70 percent savings Steinberger is currently leaving on the table will eventually be the baseline, and the baseline will be cheaper still.
At that point, the question of whether a three-person team running 100 agents is efficient will seem quaint. The answer, it turns out, was always yes.