# Tracer

- **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:** [Zuhriddin Sharofiddinov](https://cerebralvalley.ai/u/zukhriddin_ds), [Najmiddin Mukhitdinov](https://cerebralvalley.ai/u/najmitdinm), [Amirkhan Abay](https://cerebralvalley.ai/u/agpoev)
- **GitHub:** https://github.com/Zuck7/tracer
- **Demo video:** https://youtu.be/y0ZbhBL7NsI
- **Gallery:** https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/75

The problem: pulling a product from the market, whether for contamination, an undeclared allergen or a labelling mistake, means acting across warehouse, order, carrier and customer systems that don't talk to each other. Today people coordinate it with spreadsheets and phone calls over days, while unsafe product keeps shipping. When automation fails halfway, nobody knows for sure which units are on hold, which orders were cancelled, or which customers were warned once, twice or never. The result is people at risk, fines and lawsuits, lost trust, and no clean record for regulators. AI agents could do this work, but most stop at "200 OK", which isn't proof the job is done.

Tracer is an agent that runs a recall across all of these systems and doesn't call the job finished until every step is proven in the real system. From a plain-language report by the supplier, it finds every affected unit, puts stock on hold, cancels open orders, intercepts shipments on the road, notifies each customer exactly once, and produces a report where every line carries proof.

It keeps going when things break:
- A system says "200 OK" but did nothing: Tracer reads every record back before counting a step as done.
- The agent crashes mid-run: it restarts, checks what really happened, and continues without skipping or repeating anything.
- A partner changes its API: Claude writes a fix, Tracer tests it on one shipment, uses it for the rest, and saves it so the next recall fixes itself with no AI call.
- The world changes: if a truck leaves before the cancel lands, Tracer intercepts the shipment instead.

Impact: days of manual coordination become one supervised run, with no unit missed, no customer notified twice, and an audit trail ready for regulators. In our test across 15,000 orders and 13,000 shipments, with every failure triggered on purpose, all 85 steps across two recalls were verified and every customer got exactly one notice.

The same approach works for any warehouse job that touches several systems and must be finished correctly: stock holds, mass order cancellations, refunds, supplier switches. A human stays in charge: the supplier approves the plan before anything changes and can pause the agent at any time. All data is synthetic.

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

- [Swarmtect](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/72)
- [HandOver](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/73)
- [Customer Onboarding : Agents in Action](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/74)
- [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)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/75. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt