Agent online
Recursive Agents Hackathon · 2026-10-17 · Crusoe Managed Inference · NVIDIA Nemotron

Supplier Chief of Staff. Agents that keep a pharma plant from running out.

A sentinel watches every material at every site. When stock runs low, the agent writes its own definition of done, spawns sub-agents to find a transfer, negotiate with approved suppliers and scout the web, has an independent judge check the result, and texts the planner only when a decision is above its authority. It learns supplier reliability from every delivery.

Watch a mission live → Message @supplychief_bot Evidence board

Corvane Biologics is a fictional company; its inventory, suppliers and prices are synthetic. The models, orchestration, checks and approvals are real.

Stockout events−79%

vs a manual planner, 90-day simulation, mean of 10 seeds

Planner minutes−99%

21,450 → 228 over 90 days; mostly one-tap approvals

Learning−22%

stockout days vs the rule-based agent when suppliers overstate reliability

Production mission6/6

checks passed, about 3.5 minutes end to end on Vercel

How a mission runs

01 · code

Sentinel

Scans every position daily with no model call. A breach opens a mission.

02 · definition of done

Six checks first

Covers the need, lands in time, qualified suppliers only, comparable quotes, batches stay feasible, every claim grounded.

03 · recursion

Sub-agents fan out

Transfer, sourcing and scout in parallel. Sourcing spawns one negotiator per approved supplier.

04 · judge

Verify, re-plan or ask

Code recomputes every number; a different model checks the claims. Up to 3 attempts, then a human above $25k.

Recursive agent tree: planner, run_mission workflow, transfer, sourcing with negotiators, scout, and an independent verifier
The agent tree. The judge never shares a model with the doers. Full drawing set: architecture.pdf.

Go deeper

Live

Mission theater

A simulated phone wired to this agent, next to the pipeline driven by its real event stream.

 Live

Evidence board

Structure images, line-feasibility tables, QA packets and supplier evidence from missions.

 Code

GitHub

Agent app, shared core, simulator, evals. Everything built during the hackathon.

 Results

Self-improvement

Supplier reliability learned from deliveries, ranked by risk-adjusted cost per client.

 Results

Simulation

Manual planner vs agent on the same demand and supplier delays.

 Results

Model evals

Tool selection, guardrails, spec matching and negotiation across four Nemotron models.

Talk to it over HTTP

curl -s -X POST https://supplier-chief-of-staff.vercel.app/eve/v1/session \
  -H 'content-type: application/json' \
  -d '{"message":"What is running low at Ghent?"}'

curl -N https://supplier-chief-of-staff.vercel.app/eve/v1/session/<sessionId>/stream