Boris Cherny on running dozens of agents at Anthropic (not a 'graph engineering' talk)

Try this

1. Give AI a goal and access to your tools, and let it coordinate the steps, instead of only asking it to finish drafts. 04:13
2. Describe the starting point and the end result you want, rather than a rigid step-by-step script. 04:29
3. Keep a standing context file (Anthropic uses CLAUDE.md) that loads your business basics into every session. 05:04
4. Set up an AI 'first stop' for new staff: where files live, how to file expenses, what the policies are. 05:48
5. Pick one bottleneck, let AI clear it, then move to the next, rather than automating everything at once. 08:34
6. Start with one agent on one task, and run several in parallel only once you trust its output. 13:58
7. For big jobs, have a main agent split the work and hand pieces to helper agents. 15:03
8. Give hard problems more effort (a stronger model or higher thinking setting) instead of expecting instant answers. 18:24
9. Judge AI by the experiments it lets your team try, not only by time saved. 19:42

Timestamps point to the moment in the source video.