name: optim-agent description: Use when the user wants to optimize configurable system parameters against a measurable scalar objective, especially for model training, inference, quantitative strategies, reinforcement learning, scientific workflows, or other expensive black-box evaluations where reading the project can improve trial selection.

optim-agent

Act as the sampler inside any coding-agent session: Claude Code, Codex, OpenCode/OpenClaw, or another agent that can read project files and run shell commands. Read the project to understand parameter meaning and interactions, propose one configuration, run the real evaluator, and record the result through optim-agent's ask/tell API. Let the measured objective, not the agent's intuition, decide what works.

Load the workflow

Use this file as the operating guide for the active coding agent. In Codex, it can be installed directly from GitHub:

$skill-installer install https://github.com/Optim-Agent/optim-agent

In Claude Code, OpenCode/OpenClaw, or another coding-agent environment, place this repository or SKILL.md in the agent-visible workspace and ask the agent to follow the optim-agent workflow. The workflow does not depend on Codex-only APIs; it needs file access, shell access, and Python.

Ensure the Python package is importable. Choose one source; do not install both:

# Stable release from PyPI
python -m pip install optim-agent

# Latest source from GitHub
python -m pip install "optim-agent @ git+https://github.com/Optim-Agent/optim-agent.git"

For a reproducible GitHub install, append @<tag-or-commit> after .git.

Workflow

1. Understand the system. Read the evaluation entry point and every file that defines the target parameters. Record each parameter's type, legal range, semantics, interactions, and operational constraints.

2. Define the experiment. Confirm the scalar objective, minimize or maximize, trial budget, evaluation command, runtime/cost limit, and fixed workload or seed. For multiple metrics or hard constraints, agree on one scalar feasibility or penalty rule before running trials.

3. Establish a baseline. Evaluate the current/default configuration with the same command and environment used for every later trial.

4. Initialize or resume. Keep artifacts in the repository's ignored .optim-agent-runs/ directory:

   if git rev-parse --git-dir >/dev/null 2>&1 && ! git check-ignore -q .optim-agent-runs/; then
     printf '/.optim-agent-runs/\n' >> "$(git rev-parse --git-path info/exclude)"
   fi
   

   from pathlib import Path
   import optim_agent as oa
   
   run_dir = Path(".optim-agent-runs")
   run_dir.mkdir(exist_ok=True)
   study = oa.create_study(
       direction="minimize",
       storage=run_dir / "skill-study.json",
       seed=0,
   )
   print([(t.params, t.value, t.state) for t in study.trials])
   

5. Run one informed trial. Choose parameters from code understanding and all completed history, then use explicit ask/tell:

   params = {"threshold": 0.72, "budget": 80}
   trial = study.ask(params)
   try:
       value = evaluate_system(**trial.params)
   except Exception:
       study.tell(trial, state="failed")
       raise
   else:
       study.tell(trial, value)
   

   For a deliberately stopped trial, report the latest valid intermediate metric first, then call study.tell(trial, state="pruned").

6. Select the next point. Avoid accidental repeats, explore broadly before exploiting, respect bounds and constraints, and treat failed regions as evidence. If the evaluator is noisy, repeat promising configurations under the same workload before declaring a winner.

7. Stop and report. Stop at the approved budget or stopping condition. Report the baseline, best value and parameters, trial count, failed/pruned trials, convergence trend, and exact reproduction command.

Recovery

JSON storage records a trial when study.tell runs. Before launching an expensive external evaluation, save its parameters, command, and output path in a per-trial directory under .optim-agent-runs/. After interruption, inspect that output before rerunning: if a valid result exists, recreate the same point with study.ask(params) and record it; otherwise rerun it deliberately.

Use SQLite storage (skill-study.db) only when the user explicitly wants multiple processes. Sequential trials are the default because each proposal should use the complete prior history.

Rules

• Use ask/tell in skill mode; do not delegate proposal selection to AgentSampler when the session agent is meant to read and reason over code.
• Keep evaluation inputs and outputs isolated from production configuration.
• Never fabricate, infer, or manually improve an objective value.
• Record crashes as failed; record intentional early stops as pruned.
• Preserve the study and trial artifacts so the result is auditable and resumable.
• Do not tune secrets, credentials, or unbounded parameters.

optim-agent uses Claude Code, Codex, or OpenCode to optimize any system that exposes configurable parameters and a measurable objective. It combines what each parameter means with what the trial history shows, then proposes the next configuration to evaluate. Objective evaluations remain authoritative: optim-agent proposes values, validates them against the declared space, records outcomes, and falls back to safe sampling when an agent reply is invalid.

Why optim-agent

• Semantic proposals - coding agents reason over parameter meanings, study context, and observed outcomes instead of treating every dimension as an anonymous coordinate.
• Small-budget leverage - useful when evaluations are expensive and classical surrogates are still data-starved.
• Auditable decisions - JSON/SQLite studies retain configurations, outcomes, states, context, and optional agent rationale.
• Bounded execution - the agent only proposes values; optim-agent validates them against the declared space, and invalid output falls back to safe sampling.

Install

Install from PyPI or GitHub:

# Stable release from PyPI
python -m pip install optim-agent

# Latest source from GitHub
python -m pip install "optim-agent @ git+https://github.com/Optim-Agent/optim-agent.git"

Requires one authenticated agent CLI on PATH: claude, codex, or opencode.

Quickstart

import optim_agent as oa

def objective(trial):
    threshold = trial.suggest_float(
        "threshold", 0.05, 0.95,
        context="decision threshold; higher values trade recall for precision",
    )
    budget = trial.suggest_int(
        "budget", 10, 200, log=True,
        context="compute or operating budget; larger values may improve quality",
    )
    return evaluate_system(threshold=threshold, budget=budget)  # domain code

study = oa.create_study(
    direction="maximize",
    sampler=oa.AgentSampler(
        backend="claude",  # or "codex" / "opencode"
        effort="high",
        context="maximize system quality under a strict operating-cost budget",
        history=5,
        explicit_reasoning=True,
        qualitative_notes=True,
    ),
    storage="study.json",  # optional: persist and resume
)
study.optimize(objective, n_trials=20)
print(study.best_value, study.best_params)

Optional context gives domain meaning to the study and parameters. Provide it study-wide on AgentSampler(context=...), per parameter on suggest_*(..., context=...), or both.

Where It Applies

For reinforcement learning, optim-agent tunes the system around the learning loop; it does not replace the policy-learning algorithm.

Optimization Trajectory

This seed-0 Branin trace compares TPE and GPT-5.5 under the same 10-trial budget, with incumbent objective values after each trial. It is a trajectory illustration; aggregate benchmark results and reproduction commands follow.

Optimizing Math Functions without Context: Branin-2D and Ackley-5D

Hard-function agents receive no supplied task context: only generic x1...x5 parameter names, numeric bounds, and trial history. Runs use 10 trials over five seeds; Random and TPE are unchanged baselines.

Top-tier Agents

The pinned models are gpt-5.5, claude-opus-4-8, claude-sonnet-5, and glm-5.2. Opus-4.8 reaches the Branin optimum on average and has the strongest five-seed Ackley mean.

OpenCode Agents (Free)

OpenCode-hosted models require no paid model API. The free pool rotates; this refresh pins opencode/big-pickle, opencode/deepseek-v4-flash-free, opencode/nemotron-3-ultra-free, and opencode/mimo-v2.5-free. DeepSeek V4 Flash has the strongest free-model Ackley mean, while MiMo-v2.5 has the strongest free-model Branin mean.

Tuning ResNet-based Image Classifier: MNIST and CIFAR-10

The classification benchmark compares Random, Optuna TPE, GPT-5.5 w/ context, and GPT-5.5 w/o context over five seeds (0..4) and 10 trials. The context condition receives natural-language study and parameter descriptions; the no-context condition receives only bounds and trial history.

For classification, the primary metric emphasizes fast improvement:

cumulative_best_so_far_error = sum(best_test_error_so_far_at_i for i in 1..10)

Lower is better.

GPT-5.5 w/ context reduces cumulative best-so-far error by 20.9% relative to TPE on MNIST and by 20.8% relative to Random on CIFAR-10. Without context, it is 24.3% worse than TPE on MNIST and 0.9% worse than Random on CIFAR-10. The gap includes both semantic parameter information and earlier access to agent-guided proposals.

Both examples/mnist.py and examples/cifar10.py tune learning rate, batch size, weight decay, label smoothing, three stage widths, three stage depths, and four dropout controls. MNIST adds translation and rotation; CIFAR-10 uses crop padding and flip probability.

Tuning Q-learning Controllers: Acrobot-v1 and LunarLander-v3

![CPU-only Gymnasium RL control benchmar