Run limits: When should an agent stop?
A failed run can still produce useful work. Stripe, Dropbox, and DoorDash use different limits.
Lessons
Short observations for people who build agents. Each lesson examines a design choice from the agent catalog. Sources describe what teams report. Our observations explain what those reports may mean for other builders.
Human attention: When do more review comments mean more work?
Uber and HubSpot check review comments before engineers see them.
Agent roles: Should one agent search and a different one check?
DoorDash changed how its agents divide a code review. Each design exposed a different problem.
Work state: How can the work continue after the agent stops?
A new worker can continue from a saved record.
Tools and context: How should an agent find its tools?
Cloudflare and Sentry defer different parts of tool setup until a task needs them.
Workflow control: Which steps do not need a model?
Code can control a step that must follow a fixed rule.
Evaluation: How do you test an agent on your own work?
Past tasks and failures can become repeatable checks.
These lessons describe selected cases. They do not establish that one design works best for every team.
Send a public write-up, talk, or post about a company's internal agent, or tell us what to fix in an entry. Pull requests are welcome.