Hey Product Hunt,

We built Taiafox in one weekend at HackSpain 2026 for HappyRobot's challenge: can an AI actually run a crisis, not just summarize it on a dashboard?

To find out, we set a forest on fire. In Minecraft. On purpose.

A wildfire starts on the west ridge, with two villages downwind, a hospital, two fire trucks, one ambulance and a drone. Taiafox takes it all in and decides what happens next, by phone. Every call goes through HappyRobot, in and out: mayors call in and get called back, firefighters get fronts ranked by priority, the single ambulance goes where it's needed most. Residents get a calmer voice thanks to Humalike, because people hang up on robots.

Two things we like most:

It doesn't confuse guesses with facts. While a call is live, TypeSafe's jev-1.13 picks from ids that exist in the scenario every 5 seconds, so "the road by the old mill" becomes a real edge in our world model. If something's missing, Taiafox asks a follow-up like a dispatcher would, and when it runs out of budget it guesses. Every fact is tagged observed, inferred or assumed, and guesses show up grey and italic. Only observed facts can create hard constraints.

The LLM doesn't get the keys. It writes a policy. A Hungarian-algorithm solver builds the actual plan and verifiers check it. The model gets one call per replan, and only when reality drifts more than 25% from the plan or a hard constraint breaks.

Minecraft is there because debugging multi-agent systems through logs is miserable. It renders our state over RCON using only /tp, /fill and /setblock, and the LLM never touches it. Everything is event-sourced, so any run can be replayed exactly. Think VAR for disaster response.

We'd love your take:

If you work in emergency response, what would make you distrust a system like this first?
Is the Minecraft layer useful or just fun to watch? We can't tell yet.
What should we break it with next? A grid blackout is ready.

Built with HappyRobot, Humalike and TypeSafe. No villagers were harmed. Several trees were.