How I Let an AI Agent Run 100 ML Experiments Overnight on a $500 GPU

Last week I let an AI agent run 100 machine learning experiments overnight on my RTX 3070. I woke up to a 25% model improvement. Here's exactly how it works.

The Setup

The agent is built on Karpathy's autoresearch concept, powered by Claude Sonnet. It runs in a loop:

1. Propose — The agent analyzes current model performance and proposes a specific code change
2. Implement — It writes the actual Python code to modify the neural network
3. Train — The modified model trains on PubMed medical text data
4. Evaluate — Loss metrics are compared against the baseline
5. Decide — If improvement > threshold, keep the change. Otherwise, revert.
6. Repeat — Go back to step 1 with updated context

The Results

Out of 100 experiments:

• 93 failed — proposed changes made the model worse or had no effect
• 7 succeeded — measurable improvements that the agent kept
• Net result — 25% improvement in model performance

The 7% hit rate sounds low, but that's the point. Research is mostly failure. The agent runs experiments I'd never have time to try manually.

What the Agent Discovered

The 7 successful experiments included:

• Learning rate scheduling changes I wouldn't have tried
• A specific attention head configuration that improved convergence
• Batch size adjustments that were counterintuitive but worked
• Layer normalization placement that contradicted my assumptions

The Hardware

This runs on consumer hardware:

• GPU: NVIDIA RTX 3070 (8GB VRAM) — ~$500
• CPU: Standard desktop AMD Ryzen
• RAM: 32GB
• Storage: 1TB NVMe SSD

Total cost for the overnight run: about $0.50 in electricity + Claude API calls. That near-zero per-experiment cost comes from running inference locally rather than through a cloud provider. Here's what serving a live LLM from a consumer GPU actually looks like in production.

Why This Matters

The traditional ML research loop is: human thinks of experiment → human implements it → human waits for training → human evaluates → human thinks of next experiment.

Each cycle takes hours or days of human attention. My agent does it in minutes and runs 24/7.

The Code

The agent is ~300 lines of Python orchestrating:

• Claude Sonnet for reasoning and code generation
• PyTorch for training
• A simple SQLite database tracking all experiments
• Git for version control of each experiment

It's not magic. It's a loop with good prompts and clear evaluation criteria.

What I Learned

1. Autonomy requires clear metrics — the agent needs an unambiguous way to measure success
2. Failure is the feature — 93% failure rate is fine when experiments are cheap
3. Consumer hardware is enough — you don't need cloud GPUs for meaningful research
4. Overnight is the killer use case — run experiments while you sleep, review results over coffee

Try It Yourself

You need:

• A GPU (even a 3060 works)
• An API key for Claude or GPT
• A clear metric to optimize
• Patience to debug the loop

The hardest part isn't the code — it's defining what "better" means for your specific model.

Curious about commissioning something like this rather than building it yourself? Here's what custom autonomous agents actually cost in 2026.

---

Want me to build an autonomous agent for your workflow? Start a project →