motive

Answer why did the agent do that? with empirically-grounded explanations for AI agent decisions.

motive uses hierarchical occlusion saliency: mask segments of the input context, measure how much each one affects the model's decision, and surface ranked importance scores and counterfactual what-ifs. No model self-reports. No local weights required. Works with any OpenAI-compatible endpoint.

Installation

pip install motive

Quickstart

from openai import AsyncOpenAI
from motive import SaliencyEngine, Segment

engine = SaliencyEngine(
    client=AsyncOpenAI(base_url="...", api_key="..."),
    model="your-model",
)

result = engine.explain(messages=messages, segments=segments, tools=tools)
print(result.summary())

See examples/demo_saliency.py for a full working example.