# CASCADE

- **Event:** [The Recursive Agents Hackathon](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon)
- **When:** Sat, Oct 17 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Nikhil Kapu](https://cerebralvalley.ai/u/nikhilk)
- **GitHub:** https://github.com/nikhil-kapu/cascade-recursive-agents
- **Demo video:** https://youtu.be/zAiioVEIwvs
- **Gallery:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/84

CASCADE is a recursive agent team you can walk away from during a compound disaster — rain → landslides → river dams → dam bursts → floods, or quake → bridge failures → gas-line fires — all running on Crusoe Managed Inference with three tiers of NVIDIA Nemotron.

WHAT THE AGENTS DO (live, every tick): A chief on Nemotron 3 Super 120B watches a risk monitor and notifies one district agent per district at risk (Nemotron 3.5 Lightning 30B). Each district agent forecasts its own next hour, grades severity (monitoring → alert → evacuate → escalate), alerts only the emergency services that tier warrants (no public alert at low risk), plans evacuation routes, crews, shelters and airlift, and spawns its own helper agents on Nemotron 3 Nano 30B. If its risk rises a tier, the chief re-dispatches it with a stronger mandate; escalated districts phone the district emergency contact and then residents with the forecast and the plan.

DONE MEANS DONE: nothing is applied until an independent verifier (Nemotron Super) checks the plan against the simulation — routes that cross hazards, shelters inside flooded cells, over-allocated crews — plus an LLM review. Rejected plans are re-planned with the reasons attached. Live today: 8 rejections re-planned in one M6.9 run.

SHARED BRAIN: chief notes, You.com research facts and verifier lessons land in a shared memory every agent reads before planning.

VERSION N+1: after each run a reflector reads a per-seed diagnosis of three held-out scenarios and rewrites the playbook (rules + structured params); a change is kept only if the held-out score rises, and discarded proposals are never re-proposed. Measured live: Bay Area Gen 0 902 → Gen 1 1021 → Gen 2 1053; Nepal 722 → 737.

WHY IT'S HARD: a deterministic cascade engine (grid hydrology, landslide dams, fire spread, aftershocks, A* routing, population-relative scoring) shared by two regions as pure config (SF Bay Area; Nepal's Trishuli basin, anchored on the real August 2026 flood with sourced facts); concurrent planning so the cascade keeps ticking while agents think; per-model rate-limit semaphores with fallbacks; a Three.js procedural terrain where water visibly flows downstream, bridges fail, and the agent tree fans out on the map; a Delegation board that makes the recursion legible; timeline scrubbing of any incident. Public repo, everything built today. Planning simulator, not a forecast.

## More from The Recursive Agents Hackathon

- [Formloop](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/81)
- [Green-Stove Agents](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/82)
- [ARC](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/83)
- [Zen](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/85)
- [Capy](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/86)
- [The last harness](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/87)

---

Markdown version of https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/84. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt