AgentSynth

Synthetic, verified multi-step agent trajectories for fine-tuning agentic LLMs — tool-use, grounded code execution, and multi-agent collaboration — with a built-in LLM-as-Judge evaluation loop. It runs offline for free and scales up with any model.

The thesis in one line: the value isn't generating agent trajectories, it's generating ones you can trust. So verification is core, not a side feature.

Install

pip install agentsynth-ai                # core: generate + evaluate + export
pip install "agentsynth-ai[app]"         # + the Gradio UI
pip install "agentsynth-ai[all]"         # everything

60-second tour

from agentsynth import AgentTrajectoryGenerator, TrajectoryEvaluator, verify_trajectory

gen = AgentTrajectoryGenerator()                      # offline mock by default
traj = gen.generate("What's the weather in Paris, and 18% tip on $54?")

result = TrajectoryEvaluator().evaluate(traj)         # 6-dimension rubric
print(result.overall, result.passed)

print(verify_trajectory(traj).verified)               # re-checks tool args, execution, safety

Where to go next

• Vision — the problem, the bet, and the principles.
• Architecture — how the pieces fit together.
• Fine-tune & benchmark — turn the data into a model and prove it helps.
• API reference — the public surface.

The live demo runs on Hugging Face Spaces, and the code is on GitHub.