Open source · Future AGI · 2025 · Python · LiteLLM · Pydantic
Agent Optimizer
An open-source library that searches for better prompts and agent workflows, scored by whatever evaluator you plug in.
What I did
I built the working prototype and the Bayesian-search optimizer, simplified the optimisation loop, and merged the first release to main in October 2025.
Who I made it with
azain-commits wrote most of the other optimizers (ProTeGi, meta-prompt, GEPA and PromptWizard) and the evaluator integration.
What it’s for
Hand-tuning a prompt stops scaling after a few examples. agent-opt treats it as search: generate candidates, score them with an evaluator, keep what improves.
It ships several strategies (random search, ProTeGi, meta-prompting, GEPA, Bayesian search and PromptWizard), and any Future AGI evaluation can be the score.
Why it was hardPrompt optimizers each assume something different about data and scoring. Putting several behind one interface meant agreeing on what a candidate, a score and a dataset are.
Receipts
- future-agi/agent-opt PR #6 — first release to main (merged 2 Oct 2025)Optimizers shipped (random search, ProTeGi, meta-prompt, GEPA, Bayesian, PromptWizard). Commit authors: Kartik (working prototype, simplified loop, Bayesian search) and azain-commits (most other optimizers, evaluator integration).
Next, a little further back in timeTriggers of shared regulation →