Critiqor: Runtime Intelligence Platform for AI Agents

Critiqor evaluates AI agents by observing their runtime behaviour, not their self-reports. Diagnose failures, build trust, and deploy with confidence.

Critiqor is a runtime reliability intelligence platform built for OpenClaw agents. Instead of asking an agent to evaluate itself, Critiqor silently observes every tool call, provider request, memory event, and state transition during execution — then diagnoses what actually happened.

Why Critiqor?

Most evaluation tools ask the agent to grade its own output. That works until it doesn’t: hallucinated tool calls look fine in the response, infinite loops produce confident-sounding answers, and memory failures are invisible unless you were watching the trace.

Critiqor solves five problems that agent self-reports cannot:

Problem	What goes wrong without Critiqor
Unreliable evaluation	Agents grade their own output — high confidence despite clear failures
No runtime visibility	Tool calls, retries, and memory events happen invisibly
Weak explainability	You know the output is wrong; you don’t know why
No root cause analysis	Failures are observed but not traced to a cause
No evidence-backed diagnosis	Guesses replace data when debugging agents

How It Works

Developer → Agent → Runtime Events → Evidence Collection → Diagnosis Engine → Dashboard → Recommendations

When you run critiqor monitor openclaw, Critiqor:

1. Attaches the bundled OpenClaw plugin before the agent starts
2. Collects every observable runtime event into a session file
3. After the session ends, runs the diagnosis engine over the evidence
4. Writes a diagnosis file and opens a local dashboard with findings

Nothing is sent to a remote server. All evidence stays on your machine.

What Critiqor Evaluates

Critiqor scores six reliability dimensions drawn from the OpenClaw failure taxonomy:

Loop Control

Detects repeated tool calls with identical arguments — infinite tool loops

Memory Integrity

Flags failed recall, ignored memory, and lost context events

Tool Output Utilization

Identifies tool outputs that were available but unused in decisions

Context Health

Catches context saturation and compaction that can hide important state

Cost Efficiency

Measures token waste and redundant execution overhead

Skill Adherence

Detects mis-selected, ignored, or failed OpenClaw skill invocations

Quick Navigation

Installation

Install Critiqor with pip and get your environment ready

Quick Start

Run your first monitored session end-to-end in four steps

CLI Reference

Full reference for monitor, finalize, dashboard, and runs commands

Dashboard

Understand diagnosis sections and how to switch between runs

Evidence Types

Deep dive into every evidence category Critiqor collects

OpenClaw Integration

Full setup guide for the OpenClaw plugin