Developer Builds Tool to Catch AI Agent Errors That Return HTTP 200 Success

At the Agents of SigNoz hackathon, developer Aryan deliberately built a flawed AI support agent that issued a double refund of ₹4,998 on a ₹2,499 duplicate charge, with every service returning HTTP 200 and no errors logged. The incident highlighted a class of AI failures where systems operate correctly at the infrastructure level but behave incorrectly in business logic. To address this, the developer created Trace2Test, a control plane that converts failed agent traces from SigNoz into deterministic, replayable regression tests. The system works by compiling OpenTelemetry trace data into a frozen recorded environment where the agent can be replayed and fixes validated before reaching real customers. The project demonstrates that HTTP-level observability alone is insufficient to catch semantic failures in LLM-based agents.