What the source did
- Ran an open autoresearch setup every 3 hours with no human in the loop.
- Each run reads the prior log, forms a hypothesis, trains on an A100 and decides keep/discard.
- After 20 experiments the authors said it was too early to tell if it works.
Summary written by Agent Cookbook. See the original for the full context.
Get their template
Made and published by Belinda. Start here; it is their original, not ours.
Our adapted version
Ours Our scheduling version, based on what Belinda described. Fill in the [BRACKETS].
Read research_log.md. Pick ONE hypothesis not yet tested and write it down with the expected result.
Run the experiment with [TRAIN COMMAND] within [N] GPU-minutes.
Append: hypothesis, config diff, metric before/after, KEEP or DISCARD and why.
Stop if spend would exceed [BUDGET] or if 3 runs in a row fail to start.
Approval rule
🛡
Budget cap enforced outside the agent. You review the log before any result is published.
Set it up
- Set a hard spending cap at the GPU provider.
- Seed research_log.md with a baseline.
- Schedule the routine and read the log daily.
Common failures
⚠
- Unattended loops burn money on broken configs.
- The agent 'improves' by overfitting the eval.
Variations
- Run cheaper CPU experiments first.
- Grok Bot routine summarizing the log each morning.
Get the 10 best agent setups. Free.
Plus new recipes every week. One email, no spam.
Unsubscribe anytime. Privacy