I’m building a circuit breaker for AI agents

I'm building a circuit breaker for AI agents

AI agents have a weird failure mode that traditional software doesn't really have.

They can be technically running successfully while doing absolutely nothing useful.

An agent might:

• Repeatedly retry a failing API
• Use an expensive model when a cheap one would work
• Get stuck in a loop that quietly burns through your API budget

So I started building Moven

Moven sits between your agent and its tools/models and watches what happens while the agent is running. When something looks wrong, it can intervene.

For example:

Agent
  ↓
search()
  ↓
search()
  ↓
search()
  ↓
search()
  ↓
🚨 MOVEN
Repeated tool call detected.
Agent paused.

Instead of discovering the problem 30 minutes later when you look at your logs, Moven can stop the run while it's happening.

Right now I'm experimenting with

• Repeat-call interception — detect agents repeatedly calling the same tool
• Hard cost ceilings — stop a run once it reaches a defined budget
• Model fallback/recovery — switch models/providers when something fails
• Automatic recovery — attempt to recover from certain failure patterns
• Runtime visibility — understand why an agent was stopped

The bigger idea

Agents shouldn't have unlimited permission to run forever. We've had circuit breakers, rate limits, timeouts, and resource limits for traditional software forever. I think agents need the same kind of runtime protection layer.

I'm calling it Moven.

---

The project is still early, and I'm trying to figure out whether this is actually a problem developers will care enough about to adopt.

If you're building AI agents: what's the worst failure you've had?

Was it:

• An infinite loop?
• An agent repeatedly calling a tool?
• A huge API bill?
• A model getting stuck?
• An agent doing something completely unexpected?

I'd genuinely love to hear the failure stories.

Moven: https://moven.dev