A 200-line bash orchestrator wires a meta-harness optimization loop onto Islo sandboxes, going 0/5 to 5/5 on a toy task suite in four proposer steps.
Key Takeaways
Core primitives: islo snapshot save for reproducible envs, islo use --snapshot for parallel candidate forks, islo logs for durable execution traces.
Bottleneck framing from Yoonho Lee: most optimizers compress runs into summary stats; meta-harness gives the proposer up to 10M tokens of raw traces to read instead.
A free transfer-fix emerged at iteration 1: a fizzbuzz hint containing “inclusive” incidentally fixed sum-evens too, visible only because the proposer reads all traces.
Swapping from the offline simulator to a real Claude/Islo backend requires three line changes; the orchestrator stub is already in place.
Next steps: replace toy tasks with Harbor evals, swap proposer.py pattern-match for Claude reading the full runs/ tree from inside a sandbox.
