Skip to main content

AgentUQ Docs

Docs v0.1.0

Runtime reliability gates for LLM agents

AgentUQ turns provider-native token logprobs into localized runtime actions 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.

The documentation in this site is the web-first presentation layer. Canonical docs stay in this repo under docs/, while the Docusaurus app lives in website/.

AgentUQ is open source under the MIT License.

  • Localized risk, not one opaque score Catch brittle SQL clauses, tool arguments, selectors, URLs, paths, shell flags, and JSON leaves before execution.

  • Probability honesty over fake certainty Use canonical scoring only for genuinely greedy runs and realized scoring for the path that actually came out.

  • Cheap first-pass gating Put a lightweight reliability layer in front of slower systems such as retrieval, semantic verification, sandboxing, or human review.

Where to start

What decisions AgentUQ can trigger

  • continue: proceed normally.
  • continue_with_annotation: proceed, but attach the result to logs, traces, or monitoring.
  • regenerate_segment: retry only the risky leaf or clause when structured repair is available.
  • retry_step / retry_step_with_constraints: rerun the whole model step.
  • dry_run_verify: run a safe validator before execution.
  • ask_user_confirmation: pause before a side effect.
  • block_execution: fail closed before anything external happens.
  • custom, emit_webhook, and escalate_to_human: dispatch through your own policy layer.

For concrete routing patterns, see Acting on decisions. For the statistical framing, read Research grounding.

Documentation map

Read this site as a DX layer, not as a separate source of truth

The rendered docs optimize for scanning, navigation, and professionalism on the web. The underlying Markdown files in the repo remain canonical, auditable, and versionable.