cmpnd and a Carnegie Mellon intern added Flex to DSPy, a component that lets the GEPA optimizer rewrite a program's own source code as well as its prompt wording. On a place-matching task it raised accuracy while cutting cost and latency, largely by handling easy cases in plain Python and calling the model only for ambiguous ones. Model-written code runs in a sandbox rather than in the host process, and a separate test on a software-fixing benchmark is described as a pilot.
What changed
DSPy optimizers could only rewrite a program's instructions, not its code.
What it unlocks
Letting an optimizer restructure a program's own Python — adding helper functions and routing easy cases away from the model — scored against a chosen cost and accuracy metric.
- 90.4% to 95.0% accuracy
- $0.98 to $0.70 per 1k records
- 75% fewer model calls
- 4 of 12 SWE-bench Pro issues fixed
What you need to act on it
- DSPy
- an optimizer run with a strong reflection model
- a labeled training set and metric
Sources