agentuq

AgentUQ

Single-pass runtime reliability gate for LLM agents using token logprobs.

AgentUQ turns provider-native token logprobs into localized runtime decisions for agent steps. It does not claim to know whether an output is true. It tells you where a generation looked brittle or ambiguous and whether the workflow should continue, annotate the trace, regenerate a risky span, retry the step, dry-run verify, ask for confirmation, or block execution.

Why teams use it

• Catch brittle action-bearing spans before execution: SQL clauses, tool arguments, selectors, URLs, paths, shell flags, and JSON leaves
• Localize risk to the exact span that matters instead of treating the whole response as one opaque score
• Spend expensive verification selectively by using AgentUQ as the first-pass gate

Install

pip install agentuq

For the OpenAI example below, also install the provider SDK:

pip install openai

For local development and contributions:

python -m venv .venv
. .venv/bin/activate
pip install -e .[dev]

Examples below assume the public package and import namespace agentuq.

Integration status

OpenAI Responses API is the stable integration path in the current docs. Every other documented provider, gateway, and framework integration is preview, including OpenAI Chat Completions, OpenRouter, LiteLLM, Gemini, Fireworks, Together, LangChain, LangGraph, and the OpenAI Agents SDK.

Minimal loop

from openai import OpenAI

from agentuq import Analyzer, UQConfig
from agentuq.adapters.openai_responses import OpenAIResponsesAdapter

client = OpenAI()
response = client.responses.create(
    model="gpt-4.1-mini",
    input="Return the single word Paris.",
    include=["message.output_text.logprobs"],
    top_logprobs=5,
    temperature=0.0,
    top_p=1.0,
)

adapter = OpenAIResponsesAdapter()
analyzer = Analyzer(UQConfig(policy="balanced", tolerance="strict"))
record = adapter.capture(
    response,
    {
        "model": "gpt-4.1-mini",
        "include": ["message.output_text.logprobs"],
        "top_logprobs": 5,
        "temperature": 0.0,
        "top_p": 1.0,
    },
)
result = analyzer.analyze_step(
    record,
    adapter.capability_report(
        response,
        {
            "model": "gpt-4.1-mini",
            "include": ["message.output_text.logprobs"],
            "top_logprobs": 5,
            "temperature": 0.0,
            "top_p": 1.0,
        },
    ),
)

print(result.pretty())

Documentation

The web docs are built with Docusaurus from the canonical Markdown in docs/ and the site app in website/.

• Start here: docs/index.mdx
• Get started: docs/get-started/index.md
• Provider and framework quickstarts: docs/quickstarts/index.md
• Concepts: docs/concepts/index.md
• API reference: docs/concepts/public_api.md
• Maintainers: docs/maintainers/index.md
• Contributing: CONTRIBUTING.md

Repo layout

• src/agentuq: library code
• examples: usage examples
• tests: offline, contract, and optional live tests
• docs: canonical documentation content
• website: Docusaurus site and Vercel-facing app

Testing

Default pytest runs only offline tests:

python -m pytest

Live smoke checks are manual and opt-in:

AGENTUQ_RUN_LIVE=1 python -m pytest -m live