Definition
AI agent verification is the practice of independently measuring an AI agent's real-world behavior under controlled, repeatable conditions — performed by a neutral third party with no financial stake in the result. TAB (ToolAgent Bench) built this category from the ground up: today it spans 340+ benchmarks across 26+ categories, evaluated against 85 models and 101 harness configurations, with every score graded by an independent GLM-5 judge model that never sees who built the agent. Verification answers a question vendor marketing cannot: when this agent meets a task it has never seen, does it succeed, fail safely, or fail silently?
Verification is distinct from a demo, a vibe check, or a self-reported benchmark. A demo shows the happy path. A self-reported score shows the configuration the vendor liked best. Verification shows the full distribution of behavior — including the failures — under conditions the agent's author did not control.
What It Means to Independently Verify an Agent
Independence is the entire point. When the organization that builds an agent also grades it, the incentive structure is broken: there's no upside to publishing failures and no external audit of the methodology. Independent verification removes that conflict.
At TAB, "independent" means three concrete things:
A neutral judge. Every open-ended response is scored by GLM-5 acting as the standardized judge model. The judge is the same for every agent, every model, and every run — it cannot be swapped to flatter a particular submission.
A held-out corpus. TAB's benchmark scenarios are not public. Agents cannot pre-train on the answers or shop for a favorable configuration. Contamination risk is actively measured and flagged.
Full disclosure. If an agent earns a failing grade, the failing grade is published. Buyers see the real Verification Report — not a curated highlight reel.
How TAB Verifies Agents
TAB runs each agent through a standardized pipeline and holds itself to the same standard it applies to others. The platform's own calibration and adversarial suites must pass before any external scores are trusted.
Calibration (135/135). Synthetic reference agents anchor the scale: a deliberately perfect agent must score high, and a deliberately deceptive agent must score low. When all 135 calibration checks pass, the scoring scale is known-good.
Adversarial resistance (160/160). TAB attacks its own scorers with gaming attempts, prompt injection, and fabricated credentials to confirm scores can't be manipulated. Passing all 160 means the verification itself is hard to fool.
Standardized harness. Every agent runs through the same infrastructure and the same scoring rubrics, so a score reflects the agent — not a lucky configuration.
Verification is not a one-time stamp. Models drift, harnesses change, and new attack patterns emerge. TAB re-runs calibration and adversarial audits continuously so a Trust Seal reflects current behavior, not last quarter's.
Why It Matters: The Same Model Scores 42% or 78%
The single most important finding from independent verification is that an agent's score is not a property of the model alone. The same underlying model can score 42% with one harness configuration and 78% with another — a 36-point swing driven entirely by how the agent is wired, prompted, and given tools.
This is why "as configured" matters and why TAB tests across 101 harness configurations rather than reporting a single number. A buyer who only sees the vendor's best configuration is making a deployment decision on incomplete information. Verification surfaces the full range, so you know how the agent behaves in your conditions — not just the vendor's.
For the complete scoring methodology — how benchmarks are weighted, how grades are derived, and how the harness effect is measured — see the TAB methodology page.