# supplier-chief-of-staff

- **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:** [Paritosh Kulk](https://cerebralvalley.ai/u/pkul)
- **GitHub:** https://github.com/paritoshk/supply_chain_recursive_agent
- **Demo video:** https://www.loom.com/share/9393ed602b2142c1bf8a2c22e50f90ed
- **Gallery:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/1

Supplier Chief of Staff is a recursive agent team that prevents a specialty-pharma manufacturer from running out of materials and asks a planner on Telegram only when a decision exceeds its authority. The company (Corvane Biologics) is fictional, and the data is synthetic; the models, orchestration, checks, and approvals are real. We spoke with a client during a hackathon who is facing this problem, and we are a funded preseed startup tackling this at getachilles.ai 

Once a day, a sentinel scans every material at every site in code. When stock falls below its reorder point, the planner opens a mission. A workflow first writes six checks for “done” in code, then runs three sub-agents in parallel: transfer, sourcing, and scout. Sourcing spawns one negotiator per approved supplier, and each negotiator haggles with a simulated supplier. A two-part judge then grades the plan. Code recomputes every number, and a different Nemotron model checks the doers’ claims against those numbers. If a check fails, the agent replans with concrete feedback up to 3 times, with a guard against repeating the same plan.

It also gets better over time. Every recorded delivery updates a probabilistic belief about that supplier. Suppliers are ranked by risk-adjusted cost for the specific client, with no fixed weights. Judge findings become lessons for the next mission. In a 360-day simulation where a quarter of suppliers overstate their reliability, the learning agent has 22% fewer stockout days than our rule-based version. Learning never changes the guardrails: approved suppliers only, no material substitutions, and a human approves anything over $25k.

## More from The Recursive Agents Hackathon

- [icomppower.com](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/2)
- [Mato](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/3)
- [Traceback](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/4)
- [The A team](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/107)
- [AutoRoboHarness](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/108)
- [Swarm](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/109)

---

Markdown version of https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/1. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt