What it does
Built for the OpenEnv Hackathon (March 2026). The agent acts as a Staffing Agency CEO managing multiple clients (demand) and candidates (supply), balancing bench costs, candidate patience, project deadlines and client satisfaction. Inspired by VendingBench: instead of stocking vending machines, the agent sources and places developers into multi-role projects.
- Starts with $50,000 seed capital, 3 clients, a pool of 20 candidates, 52 weekly steps
- Multi-role projects only bill once every role is filled; projects expire in 4–10 weeks
- Client churn (satisfaction < 0.3) costs a $50,000 LTV penalty
- OpenEnv-native (MCPEnvironment + FastMCP), deployable to Hugging Face Spaces, with a live config hot-patch API
How it works
Observe (GET tools) → Decide (LLM emits JSON tool call) → Act (EXECUTE tools) → env.step() reward → advance_week
- 19 tools: free GET tools to observe state, EXECUTE tools with real economic consequences (interview −$500, hire −$2,000, severance, matching).
- LLM judges score hidden candidate skills, fit, salary negotiation and client satisfaction.
- Hybrid reward: dense per-tool rewards, delayed rewards on
advance_week(margins, bench burn, expiry, churn), sparse $50,000 win bonus above $200,000 profit. - Training: REINFORCE/GRPO-style with discounted returns (γ = 0.99), KL penalty 0.05 vs the reference model, LoRA adapters (rank 16, alpha 32).
Key results
| Model | Reported result |
|---|---|
| Qwen3-0.6B | Mean profit −$7,803 · max +$80,760 · 50% positive-profit rate |
| Qwen3-8B | Mean profit (first episode) +$61,329 · total reward +$57,781 · 0 parse failures |
As reported in the repository README.
Tech stack
PythonOpenEnvFastMCPDockerHugging Face SpacesQwen3LoRAWeights & Biases