I built an AI incident responder that refuses to fix anything without asking
Built for the WeMakeDevs × TrueFoundry Agent Harness Hackathon. There are two kinds of "AI for incident response," and both of them are wrong. The first acts on its own. It sees latency spike, decides it knows why, and rolls back your deploy at 3am. When it's right, it's magic. When it's wrong — and it will be wrong, because production is where confident reasoning goes to die — you now have two…
Two types of AI for incident response were examined during the WeMakeDevs × TrueFoundry Agent Harness Hackathon. The first type independently takes action based on detected issues, which can lead to problems if it's incorrect. The second type simply notifies a human and provides log summaries, though this approach is both safe and ineffective. The goal was to create an AI incident responder that combines the speed of the first type with the safety of the second. The resulting solution is called Mayday.
Mayday is designed to prevent an AI from making changes without explicit human approval. When an alert is triggered, Mayday automatically scopes the incident, runs parallel analyses on metrics and logs, correlates with deployment history, and writes a diagnostic. It then proposes a single proposed fix, including root cause, expected impact, and what was ruled out.
The agent stops after proposing the fix, waiting for human approval before implementing any changes. A 401 HTTP error demonstrates that attempting to bypass this approval process is unsuccessful.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
