Optim-Agent/optim-agent
LLM agents as your hyperparameter optimizer.
What it solves
Optim-agent automates the iterative process of tuning system parameters (hyperparameter optimization) by replacing traditional mathematical surrogates with coding agents. It solves the problem of "blind" tuning where parameters are treated as anonymous coordinates, allowing agents to reason over the semantic meaning of parameters and the trial history to propose better configurations more efficiently, especially when evaluation budgets are small.
How it works
The system integrates with agent CLIs (like Claude Code, Codex, or OpenCode) to create an optimization loop. The user defines an objective function and a parameter space (with optional natural-language context for each parameter). The agent analyzes the current trial history and the provided context to propose the next set of values. To ensure stability, optim-agent validates all agent proposals against the declared parameter space and falls back to safe sampling if the agent provides an invalid response.
Who it’s for
It is designed for algorithm engineers and researchers working with systems that have configurable parameters and measurable objectives, including those in model training, inference serving, quantitative research, reinforcement learning, and scientific workflows.
Highlights
- Semantic Proposals: Uses LLMs to reason about what parameters actually mean rather than treating them as numbers.
- Agent-Driven Pruning: Includes an agent-based pruner that can decide whether to keep or stop underperforming trials based on learning curves.
- Flexible Backends: Supports multiple agent CLIs including Claude, Codex, and OpenCode.
- Distributed Execution: Supports concurrency via thread pools or shared SQLite storage for multi-process/multi-machine optimization.
- Auditable History: Stores configurations, outcomes, and agent rationales in JSON or SQLite for full transparency.