zeph-experiments

Experiment engine for adaptive agent behavior — autonomous hyperparameter search and A/B testing for Zeph.

Overview

Provides a self-experimentation loop that mutates agent configuration parameters (temperature, top-p, system prompt), runs benchmark evaluations using LLM-as-judge scoring, and tracks results in SQLite. Three search strategies — exhaustive grid sweep, random sampling, and local neighborhood search — are available. Experiments can be triggered on-demand via slash commands or scheduled via cron.

Key types

Type	Description
`Variation`	A config mutation (temperature, top-p, top-k, frequency/presence penalty, system prompt)
`ExperimentResult`	Single experiment outcome with LLM-as-judge score and latency
`ExperimentEngine`	Orchestrator: evaluates a baseline, iterates variations (greedy hill-climbing), and persists results; yields an `ExperimentSessionReport`
`BenchmarkSet` / `BenchmarkCase`	Evaluation dataset loaded from a TOML file
`Evaluator`	Runs benchmark cases with parallel judge scoring and token budget enforcement
`EvalReport`	Summary with mean score, p50/p95 latency, error count
`SearchSpace`	Parameter ranges and bounds for variation generation
`VariationGenerator`	Strategy trait — `GridStep`, `Random`, `Neighborhood` implementations
`ConfigSnapshot`	Captures the current baseline config for rollback

Usage

Experiments are launched from the agent chat via slash commands:

/experiment start         # run up to max_experiments from config
/experiment start 5       # run at most 5 experiments
/experiment stop          # cancel the running session
/experiment status        # show current session progress
/experiment report        # print results table
/experiment best          # show the top-scoring variation

[!NOTE] Only one experiment session can be active at a time. Use /experiment stop to cancel before starting a new one.

Configuration

[experiments]
enabled = true
max_experiments = 10
max_wall_time_secs = 300
eval_budget_tokens = 4096
min_improvement = 0.05       # minimum score gain to accept a variation
eval_provider = "fast"       # references a [[llm.providers]] name; empty = primary provider

# Scheduled experiments via cron
[experiments.schedule]
cron = "0 0 2 * * *"         # daily at 02:00
max_experiments_per_run = 3
max_wall_time_secs = 600

Benchmark datasets are loaded from TOML files:

# benchmark.toml
[[cases]]
prompt = "Explain Rust ownership in one sentence."
reference = "Each value has a single owner; the value is dropped when the owner goes out of scope."
tags = ["rust", "concepts"]

[[cases]]
prompt = "Write a hello world in Python."
reference = "print(\"hello world\")"

Pass the benchmark file via config:

[experiments]
benchmark_file = ".zeph/benchmarks/core.toml"

Installation

cargo add zeph-experiments

Or reference it from another workspace crate:

[dependencies]
zeph-experiments = { workspace = true }

This crate exposes no Cargo features — the engine is always available once the dependency is added.

Documentation

Full documentation: https://bug-ops.github.io/zeph/

License

MIT

zeph-experiments 0.22.0