Agentic Evals is the evaluation service of Agentic Thinking. If you build an AI agent runtime, or a product where an agent takes actions through tools, APIs or shells, we run it, try to break its guardrails, and tell you plainly what holds and what doesn't, with evidence. Every evaluation is led by a person, not a script.
Agentic Evals uses the methods from our incident research: every evaluation runs from a fixed commit, in an isolated environment, with negative controls such as tampered evidence, replayed approvals and bypass attempts. The rules that keep an evaluation independent of its fee are on our Trust page.
AgenticBench is separate: AgenticBench is separate from Agentic Evals. Harness makers can pay for private runs of the same suite before a release (Harness Watch). Payment never changes a public result: the suite and scoring are the same, and public results are published whether or not a maker pays.