Share "strategies" that are learned

[petermchale]

Currently strategies are stored locally, it seems:

https://github.com/codelion/optillm/tree/main/optillm/plugins/spl#data-storage

But some strategies may be useful to other users of optillm. Rather than have them "reinvent the wheel (strategy)", wouldn't it be nice to offer the ability to share strategies, e.g., by serving the learned strategy over MCP?

[codelion]

The strategies learnt from the experiment are here - https://github.com/codelion/optillm/blob/main/optillm/plugins/spl/data/strategies.json the idea for using MCP sounds interesting but we haven't explored it yet.

[petermchale]

Oh, I'm not interested in the strategies learned for those specific experiments.

My point is that if frameworks like optillm become mainstream in the workforce, and workers start to port their domain expertise into artifacts like optillm's strategies while using their agents, then other users in those same domains will be interested in downloading those learned strategies to their own agents, so that that they don't need to reinvent the wheel, or so they can immediately learn from others' mistakes, etc. This point is not unique to optillm; there are many frameworks in which artifacts similar to optillm's strategies are learned. Here's a non-exhaustive list:

• case bank of case-based-reasoning DS-agent (https://github.com/guosyjlu/DS-Agent)
• lessons of gptme (https://github.com/gptme/gptme-agent-template/tree/master/lessons; https://x.com/ErikBjare/status/1921483296388620557)
• archive of ADAS (https://x.com/jeffclune/status/1825551351746867502)
• tool dataset of atlass (https://arxiv.org/pdf/2503.10071)
• skill library of Voyager (Fig 2 of https://arxiv.org/pdf/2305.16291)
• archive of the AI CUDA engineer (https://pub.sakana.ai/ai-cuda-engineer/)
• global library of Ndea (https://youtu.be/5QcCeSsNRks?si=bQV2hIEibMBYn_l6&t=1917)
• MCP box of Alita (https://github.com/CharlesQ9/Alita?tab=readme-ov-file)
• program database of FunSearch (https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/), AlphaEvolve (https://arxiv.org/pdf/2506.13131), OpenEvolve (https://github.com/codelion/openevolve)

If frameworks like these move beyond research prototypes to actual use in the workplace, then marketplaces will spring up to facilitate domain-knowledge transfer via cases/strategies/lesssons/etc. Since the domain knowledge in each of these artifacts is essentially just context for an LLM, then a natural means to transfer the artifacts is Model Context Protocol.