BlackboxAI

BlackboxAI is a flight recorder and verifier for AI agents that operate software through the screen. Agents now click through CRMs, ERPs and remote desktops on our behalf, and when one reports "done" nobody can tell whether the work was really done or whether it broke a rule on the way. Give BlackboxAI a screen recording of an agent's session (any vendor, no SDK), the task the agent was given and an operating policy. It returns a step-by-step account of what the agent did, a verdict on whether the task was actually completed, the rules it broke with the frames that prove each one, a diff against earlier runs of the same task, and a sealed evidence pack an auditor can read. The policy is a plain document. Muse Spark compiles it into typed rules, each with the exact line it came from, and nothing is checked until a person reads the rules back and signs them off. The rule checks themselves are deterministic and unit-tested; only perception and the completion judgement come from models, and they are labelled as such. Runs the system is unsure about go to a review queue instead of being passed. It works where there is no DOM at all: on a legacy finance app reached over a remote desktop, it caught an agent pressing Post four times after a rejection and then claiming the invoice was posted. In our own sandbox it also records the app's accessibility events, so every step becomes a fact rather than a guess, at about a tenth of the cost. Measured on 15 demo runs: the 12 scripted runs match their specs exactly, and two real Muse Spark computer-use runs get correct verdicts. Narration recall is 0.79 and precision 0.74 over 120 ground-truth steps. Verification costs about $0.10 per short video run and about $0.01 in native mode. Live at https://blackboxai-ayudh-incs-projects.vercel.app (sign-in required).