DSPy 3.4.1 Puts Structured Reflection and Sub-Agents Behind One Budget

The change

DSPy 3.4.1, released on 2026-10-12, changes two agent-building boundaries. GEPA now defaults to InstructionProposer, which sends reflective examples through a structured dspy.Predict call and a JSONAdapter, rather than extracting an instruction from a raw formatted prompt. Separately, code executed by dspy.RLM can now import DSPy and construct Predict, ChainOfThought, ReActV2, or nested RLM sub-agents.

The release notes make the accounting and isolation rules unusually explicit: max_llm_calls covers llm_query, batched queries, and sub-agent model calls; sandboxed code may pass generation options but may not set the model, API base, API key, or headers. Only JSON crosses the interpreter boundary. In-process interpreters remain an explicit exception: they do not isolate execution, so the restrictions depend on cooperative code.

What I verified

I cloned the 3.4.1 tag at commit 099982b3c3eec9fc4ec4796d49915af42367a253 and ran Python’s compileall over the package: it passed. I also located the release’s focused RLM, facade, and GEPA test modules. Running the RLM and facade selections did not reach the tests: the repository test configuration imports litellm, which was not installed in this environment (ModuleNotFoundError). That is a test-environment failure, not a passing test result.

Why it matters

A sub-agent feature is easier to operate when recursive work consumes the same explicit call budget as the parent. Restricting credentials and routing configuration at the sandbox boundary also narrows what generated code can change. Neither control is a complete security boundary: an in-process interpreter is deliberately not isolated, and a JSON-only interface does not make model output trustworthy. The operational takeaway is to treat the interpreter choice, LM budget, and reflection model’s structured-output capability as part of the deployment configuration—not as invisible implementation details.

The upgrade has compatibility edges. GEPA’s default reflection request changes shape and now expects a dspy.LM-compatible reflection model that can return JSON; RLM prompts always include a sub-agent section; and saved 3.4.0 programs containing null Score level descriptions no longer load. Existing optimizers should therefore be re-baselined rather than compared as if the seed alone defined the experiment.