Unconference: Self-optimizing deep research agents
summary
Claims that modular, non-differentiable deep research agents can self-optimize using an LLM to critique and rewire prompts and pipelines, evaluated on ScholarQA where JEPA (genetic parareto optimization) and textgrad improve a baseline multi-module agent to match or exceed human-tuned prompts. JEPA outperforms textgrad and converges faster; starting from a simple prompt yields bigger gains than from a complex human design. System reads thousands of docs over minutes and uses LLM-as-a-judge instead of gold labels.
AI-written summary · about
chapters (8)
- 0:00Self-optimizing research agents
- 1:07Traditional RAG limitations
- 2:53How deep research agents work
- 5:03Optimizing agentic systems
- 6:05TextGrad and genetic optimization
- 8:14Experimental results
- 10:45Future directions
- 11:07Q&A
related talks
32mFrom Deep Research to Evaluation and Optimization of Advanced AI Age...
11mThe Unreasonable Effectiveness of Prompt Learning
33mAutomatic Agent Optimization and the Rise of Self-Improving Agents
01h 21mAgent Optimization with Pydantic AI: GEPA, Evals, Feedback Loops
30mIs Prompt Engineering Dead? How Auto-Optimization is Changing the Game
35mBuilding LLM-based Agents
19mTraining Agentic Reasoners
01h 57mFull Workshop: Build Your Own Deep Research Agents