# LLM deployment agent

- **Event:** [The Harness Engineering & Model Wrangling #SFTechWeek Hackathon](https://cerebralvalley.ai/e/mongodb-sftw-hackathon)
- **When:** Sun, Oct 11 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Henry Luan](https://cerebralvalley.ai/u/henryl)
- **GitHub:** https://github.com/casualcomputer/atlas-gpu-harness
- **Demo video:** https://youtu.be/WoHjJ6LMkpg
- **Gallery:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/57

Every new open model is useless until someone figures out how to run it well. Today, LLM infra engineers do that by hand, or prompt by prompt with a chat assistant that can’t run anything, can’t verify an answer, and forgets what failed.

Inference engines expose hundreds of settings (vLLM documents 180+ flags, SGLang about 135), and one latency target can involve 10,000+ configuration permutations. Good configuration is worth up to 40% on dense models and 50% on MoE models, and GPUs are scarce.

Atlas GPU Harness is an autonomous LLM deployment agent. It runs open models on real inference engines, has a coding agent propose configurations, verifies every answer, memory peak and speed, and commits each result to MongoDB Atlas before the next step. Every new proposal starts from what Atlas returns. A setting is promoted only if it beats the default on held-out questions.

Proven on real hardware:

Kimi-48B: a 30 GB model served on a 16 GB GPU. The agent read its own failure from Atlas and rewrote its code (+30.73%), but the faster engine default was kept.
Strata: a 68 GB model at 32K tokens. A +6.2% “faster” setting failed the answer check on 0 of 18 trials, so nothing was promoted.
Qwen-35B: a +3.6% “speedup” failed confirmation and was rejected.
Memory matters: with Atlas memory on, the agent needed 3 attempts instead of 4 and never repeated a known failure.
Scale: 55 experiments across 11 runs, all exported from Atlas, crash-safe and resumable.

The agent proposes, the harness decides what is true, and Atlas remembers everything. Next: vLLM and SGLang, then parallel agents tuning whole GPU fleets.

## More from The Harness Engineering & Model Wrangling #SFTechWeek Hackathon

- [By my self](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/54)
- [PaperTrail](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/55)
- [Long Horizon Quant Research](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/56)
- [LabRat](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/58)
- [Hype](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/59)
- [The Data Graduate](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/60)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/57. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt