Can AI agents run a business on their own yet?
Short answer
Not yet, outside simple things. As of mid-2026 AI keeps getting better, but you can't just say "build this" and have it done, or I wouldn't be working 12-hour days. My goal is making human headcount optional wherever AI does the work reliably, using a small team of agents I can understand and trust.
If AI agents could run the business, I wouldn't be working 12 hours a day. I'd say "build this thing" and it would just do it. We're not close, other than for simple things like a to-do list app.
People are using loops and other techniques to push in that direction, and things are getting better. They aren't at "done" yet.
What's the realistic goal?
"The goal is not 'never hire humans.' The goal is to make human headcount optional wherever capable AI can perform the work reliably."
Humans set direction and make judgment calls. AI performs as much of the digital execution as is practical: routine work, research, repetitive production, monitoring, quality checks, follow-through. People stay valuable for relationships, taste, prioritization, creativity, final evaluation and trust.
How many AI agents does a small company need?
Fewer than you'd think. For Springfield, my AI workforce, the target isn't the largest number of agents. It's a small role hierarchy I can understand and trust.
- New agents start paused.
- Each passes a controlled test run before going live.
- None goes live until its sources and tools are configured correctly.
- Shared building work goes to one dedicated builder agent, and narrower roles stay separate.
How reliable are agents in production?
Less reliable than the demos. My Slack-to-music automation went from brain dump to finished MP3 in five minutes. Then a coding agent's change broke the whole flow, it couldn't fix itself, and diagnosing it took hours. A peer who deploys agents for clients says his break daily. That's why I don't sell done-for-you agents; see projects vs. memberships.
What should "reliable" mean for a small software product?
Normally reliable, with occasional exceptions you can act on. Not a false choice between perfection and daily breakage. After a month of repeated database and page-loading incidents on one of my products, I set the standard: fix real problems, don't alert on non-problems, and handle the in-between state clearly. Constant "action required" and "blocked" emails are unacceptable.
The lesson from those incidents: define the customer journey first, and don't call a narrow repair system a solution.
Answered by TJ Larkin. From a June 2026 Local Media HQ podcast episode, peer calls in June 2026, and TJ's notes on running his agent team and product incidents in August and September 2026. Last updated .