Agent optimisation stack

15-05-2026

• ive seen various iterations, but common components in an agent optimization stack include:
  • dspy - https://github.com/stanfordnlp/dspy
    • framework for programming language models (structured and composable)
  • gepa - https://github.com/gepa-ai/gepa
    • framework for optimizing any system with textual parameters against any evaluation metric (focused on prompt variables)
    • evolutionary, uses a reflective step
  • textgrad - https://github.com/zou-group/textgrad
    • framework for building automatic differentiation via text (focused on non-prompt variables)
    • define your own loss functions and optimize them using text feedback (modelled on the pytorch api)
  • upskill - https://github.com/huggingface/upskill
    • define a problem, run agent session, use traces to generate a skill, use skill to generate test cases, run the problem through various llms (with and without the skill), iterate / improve / repeat
    • blog post on writing kernels for diffusers library - https://huggingface.co/blog/upskill
  • evoskill - https://github.com/sentient-agi/EvoSkill
    • extends the thinking behind gepa, applies it more liberally
  • skillOS - https://arxiv.org/abs/2605.06614
    • framework for skill curation for self-evolving agents
    • frozen agent executor LLM, skill curator LLM, skill repository
• key to all of the above is formalizing some task with some objective function or learning signal, and then solving for it and improving it repeatedly, “If you can measure it, you can optimize it”