# ColdTrace

- **Event:** [Agents in Context Toronto: a MongoDB Hackathon](https://cerebralvalley.ai/e/mongodb-toronto-hackathon)
- **When:** Sat, Nov 14 at 9:00 AM – 9:00 PM (EST)
- **Where:** Toronto, Canada
- **Team:** [Poneesh Kumar](https://cerebralvalley.ai/u/PoneeshK)
- **GitHub:** https://github.com/PoneeshKumar/AIC
- **Demo video:** https://www.loom.com/share/be7f3e81ab264edeaa26baf93c04682e
- **Gallery:** https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/79

ColdTrace is an agent that rescues a temperature-sensitive shipment when its flight is cancelled. A biologic that has to stay between 2 and 8 °C only has a fixed number of minutes it can spend out of range. Once a flight is lost, every minute spent waiting for a person to find a new route eats into that budget. ColdTrace handles the recovery itself. It quarantines the shipment, compares every route the carrier still offers, holds qualified cold storage for any layover, books the best route and tells the hospital the new arrival time. It never gives medical advice and never declares a product safe or unsafe. It quarantines and reroutes, and people make the final call.

The hard part is finishing the job when things go wrong halfway through. Before the agent acts, it writes each step to a ledger in MongoDB Atlas. A step only counts as done after the agent reads it back from the system it changed, so a "success" reply on its own is never trusted. Every request carries an idempotency key, and steps are claimed with atomic updates. If the agent is killed mid-run, a response is lost on the way back, or a second copy starts up, the next run checks what really happened and carries on from there. It never books twice or reserves storage twice, and the screen shows a live count of duplicate actions to prove it stays at zero.

Partner systems also change without warning. When the carrier switches to a new API format, ColdTrace stops and doesn't guess. It searches its memory of past incidents with Atlas Vector Search and Voyage AI embeddings. It then drafts a mapping and asks a person to approve or correct it. The approved rule is stored as a new memory, so the next time the same change appears the agent applies the fix straight away and keeps going.

The demo screen lets you attack the system live. You can cancel flights, kill the agent, change the carrier's API or drop a warehouse response. Then you can watch the map, the plan and the event log show the agent recovering each time.

## More from Agents in Context Toronto: a MongoDB Hackathon

- [Firewall Brigade](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/76)
- [GoalKeeper](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/77)
- [Team Nonce Sense](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/78)
- [Jovial Merryment](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/80)
- [CallE](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/81)
- [cybershah](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/82)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/79. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt