The reported OpenClaw spend is a research-scale outlier, and OpenAI reportedly covered it. Most teams will not see a seven-figure bill this week. But the mechanics are the same for a five-person engineering team using Claude Code, Codex, Cursor, or internal agents.
Agents repeatedly explore repos. They read generated files. They miss validation paths. They carry stale context. They run tool loops that look busy while burning tokens. With one agent, that is annoying. With ten agents or a shared team plan, it becomes budget and throughput risk.
Public tree scan: openclaw/openclaw
A public GitHub tree scan is not a private audit, but it shows why fleet-scale agents need repo-level controls before cost becomes visible.
- 17,933 tracked files, including 15,413 source files.
- 54 tracked blobs larger than 180 KB.
- CLAUDE.md and AGENTS.md are present.
- No `.claudeignore` was detected in the public tree scan.
What to do before scaling agents
- Give agents a compact repo map before their first turn.
- Exclude lockfiles, generated directories, snapshots, and large fixtures from casual reads.
- Write one local validation path the agent can run without guessing.
- Measure a real task before and after the fix, not just total spend.
Need this on a private repo?
The 48-hour audit finds repo and workflow cost leaks, ships a private report, adds a CI-ready threshold gate where useful, and includes one fix path so the next agent run starts cleaner.