Skip to content
Agent Radar

DSPy 3.4.1 makes GEPA's InstructionProposer the default

Updated DSPy

Summary

DSPy 3.4.1, released October 12, 2026, makes GEPA's structured InstructionProposer the default and lets dspy.RLM build DSPy sub-agents.

Tool
DSPy
Category
Frameworks
Date
Open source
Yes
License
MIT
Pricing
Not stated in the sources we cited.
Sources

DSPy 3.4.1 was published on GitHub on October 12, 2026 by the stanfordnlp/dspy project. According to the release notes, it makes GEPA's structured InstructionProposer the default instruction proposer and lets dspy.RLM build DSPy sub-agents from inside its REPL. It also adds Python 3.15 support and fixes a 3.4.0 import crash.

Release facts

FactDetail
ProductDSPy
Version3.4.1 (tag 3.4.1)
Release dateOctober 12, 2026
LicenseMIT, as reported by the GitHub API
Where to get itpip install --upgrade "dspy==3.4.1" or the GitHub release

What changed

  • New default GEPA proposer. dspy.GEPA now builds InstructionProposer() when instruction_proposer= is not passed. The release notes say it calls the reflection LM through dspy.Predict with a JSONAdapter, so history inputs, tool trajectories and image types reach the reflection LM as structured content.
  • Proposer options. InstructionProposer accepts skills, additional_instructions, base_instructions, max_chars, truncate_history_outputs and adapter.
  • RLM sub-agents. Code written by dspy.RLM can now build dspy.Predict, dspy.ChainOfThought, dspy.ReActV2 or nested dspy.RLM sub-agents. The notes say sub-agents run on the host and only JSON crosses the interpreter boundary.
  • One call budget. max_llm_calls now covers llm_query, llm_query_batched and every sub-agent LM call. Sandboxed code can only set generation parameters, not model, api_base, api_key or headers.
  • Upgrade notices. Optimization results can differ from 3.4.0 even with the same seed, and the reflection LM must be a dspy.LM that returns JSON. Programs saved on 3.4.0 with a null Score level no longer load.

Questions

Will GEPA results match 3.4.0 after upgrading? Not necessarily. The release notes say the default reflection request changes shape, so results can differ even with the same seed.

What happens if the reflection model does not return JSON? The notes say the proposal fails, GEPA logs the failure and keeps the current candidate, so a run can finish with the seed program unchanged.

Is the 3.5 LM-interface cutoff still on? Yes. The notes say legacy forward()/aforward() custom LMs, legacy-engine wrappers, complete_legacy() shortcuts and OpenAI-style messages= calls will be removed in 3.5, while lm("hello") stays supported.

See the DSPy hub, the frameworks category and the related adk-python 2.12.0 launch for more.

Alternatives

DSPy compared with its alternatives
ToolLicensePricing note
DSPyThis launchDSPy is a framework for programming, not prompting, language models, maintained in stanfordnlp/dspy.
MITNot stated
Pydantic AIPydantic AI is a model-agnostic Python framework for building typed AI agents and applications.
MITNot stated
LangGraphLangGraph is an open-source orchestration runtime for building stateful, resilient agents.
MITNot stated
MastraTypeScript agent framework. Mastra Code is the terminal agent; AgentController is the extracted loop, in @mastra/core 1.47.0 or later.
Open sourceMastra Code and AgentController are announced as open source. No license fee is stated in the posts we cited.
AgnoAgno is an orchestration and service layer for agents, exposing them through a REST API and an MCP server.
Apache-2.0Not stated

Questions

What is DSPy?

The GitHub repository describes DSPy as the framework for programming, not prompting, language models.

Is DSPy open source?

The GitHub API reports an MIT license for the stanfordnlp/dspy repository.

What is the latest launch covered?

DSPy 3.4.1, published on GitHub on October 12, 2026, makes GEPA's InstructionProposer the default and lets dspy.RLM build sub-agents.

Filed on Agent Radar. Claims in this note are limited to the sources above.