circuit breaker
A guardrail for AI finance agents. Before an agent runs an action, circuit breaker checks the risk, compares it with a threshold, and keeps a clear log of what happened. I built it at AI Engine Hackathon in Scotland.
What I did
I built the prototype at AI Engine Hackathon in Scotland.
Why I built it
AI agents can call finance tools very quickly. That is useful, but it also means a risky action can happen before a person has time to notice.
How it works
The idea is simple: put a small checkpoint before the real action. The agent proposes an action, GPT-4 gives a risk score and reason, and the system decides whether to allow, review, or block it.
What came out of it
- Showed how an agent can be checked before it touches a finance tool.
- Made the reason for each decision easy to inspect later.
- Kept the risk thresholds separate from the agent code, so the rules can change without rewriting everything.
Main pieces
- A small schema for the action the agent wants to run.
- A risk prompt that returns severity, confidence, and a short explanation.
- A decision gate that stops the action if it crosses the threshold.
What I would improve
This is not trying to be the fastest possible path. It is more useful when the action is important enough that a small review step is worth it.