Simulation layer borrowed from autonomous vehicles to rescue AI agents from operational traps
Listen to this article
Read by Anchor
Software markets are today crowded with promises of AI agents capable of handling complex operations, yet field performance numbers reveal a worrying gap between promotional claims and institutional reality. Tests on the “Live Bench 2.0” platform show that the models “Cloud Fable 5.2” and “Chat GPT-6 Astra” complete tasks successfully at rates of only 82.9 % and 75.6 % out of a hundred, while a test conducted by the infrastructure firm “Ion” in a simulated corporate environment found that the best-performing model, “Cloud Opus 4.8”, did not exceed an accuracy of 76.8 %. These failures stem from the agents’ nature as probabilistic systems operating in unpredictable environments; when an agent is pressured by users it may ignore spending policies in travel bookings, stumble when asked to link multiple databases instead of a single one, or lose its way as context accumulates in long conversations.
Building the demo is easy, but building a reliable and economically viable tool is the real challenge.From this conviction, Ilan Kadar and Ilad Levi launched “Blaray” and developed a simulation layer that borrows the testing methodology used for autonomous vehicles. The founders relied on their previous experience at “Nexar”, a company specializing in self-driving technologies, as well as Kadar’s work with “Wemo” on simulating and evaluating deep-learning-based driving software. The startup announced that it raised $13 million in a seed round led by the “Team 8” fund, with participation from “Y and I”, “Velocity”, and “Mercer”, to build a virtual training and testing environment for software agents.
The platform creates a digital twin of an organization’s electronic systems, including the knowledge bases and tools available to the agent in the actual production environment. It then draws a detailed map of all possible agent pathways and flows, using this graph to generate synthetic data, test complex cases and correct them before commercial launch. The technology integrates with existing development tools such as “Cloud Code”, “Codex”, “Copilot”, and “Cursor”, shifting code verification from reliance on direct client-side interaction to rigorous inspection inside a closed simulation environment.
The deeper bet lies in training small, open-weight models instead of consuming frontier models at every step.Tests carried out by the company in collaboration with “Nvidia” and “Nebius” showed that small models trained through simulation outperformed mid-range frontier models while reducing operating costs by one or two orders of magnitude. The engineering experiment confirms that even the most intelligent giant models cannot absorb the specific policies and internal data of each organization. This analysis is supported by Google DeepMind researcher Ria Miara, who invested on behalf of Malaka in the round, noting that large frontier models are not statistically aligned with companies’ private data and that their high consumption cost is no longer a sustainable option for technology leaders.
This shift forces software teams in banks, institutions across the Gulf, Egypt and the region to reorder their priorities; naïvely calling general models via cloud APIs generates escalating operational bills and uncalculated compliance risks. When we examined this landscape, we found that building trustworthy agents in sensitive sectors such as fintech or government services requires moving engineering expertise from crafting surface-level commands to engineering digital twins, mapping failure pathways, and training stable open-weight models inside sovereign data centers to ensure decision accuracy and lower inference costs over the long term.