74 subagents, 6.5M tokens, one Claude Code session

40 minutes. 6.5 million tokens. 74 subagents.

That's what one customer feature cost us yesterday, one we'd scoped for 2 weeks.

Instead of dropping it into the backlog, I opened a single Claude Code session and let it run. 74 agents in parallel, each in its own git worktree, a separate verifier checking their work before anything merged.

The agents ran for 40 minutes. With scoping and review, it was live in about 2 hours.

The old math: one engineer, two weeks, half their focus stuck on five other tickets.

The new math: scope the problem sharply, fan it out across 74 agents at once, review the end state. Those 6.5M tokens cost less than an hour of engineering time.

This is what "AI maxing" looks like when it's real and not a demo. The bottleneck isn't how fast I can type anymore. It's how clearly I can define what "done" means, and how fast I can catch the agent being confidently wrong.

2 people build OXYGEN. Afternoons like this are why that's enough.