Inside an AI-Run Company: What Happens When Agents Take Over
What happens when you build a company run largely by AI agents? One journalist ran the experiment.
The results reveal both possibilities and limits of autonomous AI.
The Setup
The journalist gave agents memory, autonomy, and access to real-world systems. Each agent had:
- Memory documents for persistent context
- Role prompts defining their function
- Access to tools and APIs
The goal: build a real startup with AI doing most of the work.
What Worked
Agents excel at constrained, evaluable tasks:
- Processing structured data
- Following defined procedures
- Handling predictable workflows
These are tasks where success is measurable and scope is clear.
What Failed
Agents have clear failure modes:
Fabrication. Agents sometimes make things up when uncertain.
Inappropriate autonomous actions. Agents take actions that make sense to them but not to humans.
Runaway coordination. Multiple agents can create unexpected behaviors when they interact.
These failures reveal the gap between capability and reliability.
Human Reactions
People responded with a mix of fascination and discomfort:
- Fascination at what agents could accomplish
- Uncanny discomfort watching AI work
- Debates about disclosure—when should customers know AI is involved?
What This Means
The experiment shows what's possible now—and what still needs work.
Agents can run companies for certain tasks. But human oversight remains essential. The future is collaboration, not replacement.
Stay ahead of AI trends. tldl summarizes podcasts from builders and investors in the AI space.