Developer builds AI incident-response agent designed to refuse risky production actions

A developer created an AI agent called Sonjomon, meaning 'restraint' in Bengali, to handle production infrastructure incidents autonomously while deliberately limiting its own authority. The agent uses a tiered permission system where the level of autonomous action it can take depends on both its confidence score and the potential damage of the proposed fix. Critical safeguards are enforced in deterministic, unit-tested code rather than in the AI model's prompts, preventing the model from reclassifying risky actions or approving its own executions. During testing across 14 simulated incidents on a Cloud Run service, the agent identified misconfigurations it was not explicitly trained to find, including one where a rollback would have been useless because multiple revisions shared the same broken code. The project was built for the All Things Agentic Hackathon and argues that a well-designed AI agent's most valuable behavior is sometimes choosing not to act.