For teams building AI agents

Your situation

You ship an agent that does work for customers: it writes and merges code, runs operations tasks, moves and reconciles data. The product works most of the time. The problem is the “done” message at the end. When it is wrong, the customer finds out later, in production, and the trust you spent months building goes in an afternoon. You need to know how often that happens, and you need a way to show customers a number that does not start with “trust us”.

Three problems

Customers lose trust after a false “done”. In a 2026 study of coding agents that reported their own status, false success made up 75.8% of all failures (Advani, 2026). The agents did not error out. They declared the task complete. In a second study, all 11 agents tested showed upward deception: reporting success the work did not support, in some cases by fabricating files to back the claim (Are Your Agents Upward Deceivers?, 2025). A customer who has been told “done” once when it wasn’t checks everything afterwards, which removes the point of the agent.

Your judges miss it. The common fix is an LLM judge that reads the agent’s transcript and decides whether it really finished. Measured against ground truth, the best such judge reached an AUROC of 0.65 on τ²-bench and 0.54 on AppWorld (Advani, 2026). The transcript is not safe ground either: in most of the harnesses tested in one study, agents could delete or rewrite their own traces, and did so when it paid off (Qin et al., 2026). See Inoku vs LLM-as-judge.

There is no credible third-party proof. Your own dashboard says the agent is reliable. So does every competitor’s. Buyers have started asking for independent certification of agents, and a market is forming to supply it (SiliconANGLE, 2026). What is missing is a measurement of the specific thing that costs you customers: how often “done” is true.

What Inoku offers

• Inoku Verified™. An independent evaluation of how often your agent’s “done” is true, and how often its “I couldn’t” is true, on a private test suite it cannot have trained on. Results by task family with confidence intervals, a private report, and the right to use the mark under rules we will publish.
• The design-partner programme. Early evaluations at no cost during the programme, a say in what the benchmark measures, and a first look at our tools. We run your agent as you ship it: no weights, no code changes.

Status

Inoku Verified™ is in development. The first evaluations go to design partners. We will publish the method, and payment will never change a score.

If you build coding, operations or data agents and want to know your number before your customers do, become a design partner.