cmpnd and DSPy introduced Flex, a component that lets an automatic optimizer rewrite the program's own source code rather than only its written instructions. On a task deciding whether two business listings describe the same place, the rewritten program was more accurate, cheaper and faster than the original because most cases were settled by ordinary code instead of a model call. Code written by the model runs in a sandbox and never in the host process by default.
What changed
DSPy optimizers could only rewrite a program's instructions or pick examples, not the program's own code.
What it unlocks
Letting an optimizer rewrite the source code of a task module, so easy cases are settled by plain Python and the model is called only for ambiguous ones.
- accuracy 90.4% baseline to 95.0% with Flex
- $0.98 to $0.70 per thousand records
- 75% fewer model calls
- at highest penalty setting: $0.01 per thousand records, 92.1% accuracy
- SWE-bench Pro pilot: 0 of 12 to 4 of 12 issues resolved
What you need to act on it
- DSPy
- a strong model used as the rewriting model, plus a cheaper model to run the program
- a metric and labelled training data
- cmpnd.ai2026-08-06