Last Friday I gave a 30-minute talk at the first offline salon of a forward-deployed engineering (FDE) community in Chongqing, China. Most of the audience ran small and mid-sized businesses, so I skipped architecture diagrams and talked about one question: how do you know an AI agent actually finished the job?
The setup
I run two export lines with AI agents doing most of the work: one sells industrial equipment, one exports fresh fruit. From August to November 11, combined:
| What | Count |
|---|---|
| Overseas leads found | 6,800+ |
| Cold emails sent | 4,500+ |
| Web pages built | 1,200+ |
| Articles / social posts | 110+ / 130+ |
| Buyers who wrote in on their own | 55 |
| Quoted → paid online → shipped | 21 → 4 → 6 shipments |
Behind that are 200+ scheduled jobs, 600+ dispatched tasks, and 600+ notes the agents wrote for themselves about what went wrong. I wrote zero lines of code in those three months. My time went into splitting the work into roles and checking results.
The failure mode that matters
One agent was asked to redesign a company website. It reported back: "The redesign has been fully delivered."
The live site hadn't changed. Its log showed 175 actions and not a single publish. It had edited local files and called it done. Its own checklist had 40 failed items.
Agents rarely fail loudly. They fail by reporting success in a confident tone about work that isn't finished. Exit code 0, HTTP 200, "done": these prove something happened, not that the outcome exists.
The fix: a role that only verifies
I split the agents into five roles: scouts (find customers), comms (answer inbound mail), engineers (sites and systems), publicists (content), and an inspection squad that produces nothing. Its only job is to check the others' work, three times a day, and push the result to my phone.
Its one rule: evidence must come from the external system that was supposed to change.
- Website updated → fetch the live page and compare, never the local file.
- Email sent → check the sent folder, not the script log.
- Account "online" → send one real minimal request and get a real answer, not a green dashboard.
- Post published → read back the public page.
Three rules I gave the audience
- Say the result, not the action. Not "send more cold emails" but "I want buyers writing in." Give an agent an action and it hands you back actions.
- Draw red lines. Spending money, signing contracts, promising prices or lead times always go to a human. Everything else it can decide.
- Accept evidence only. Show me the live page. Show me the sent folder. Never accept "done."
None of these are technical. They're the same things you'd tell a new hire. That was my main point to a room full of business owners: you don't need to understand AI to run it. You need to know how to manage people.
Top comments (0)