Arcade.dev has secured $60 million to enhance its enterprise AI agents.

      The issue with allowing an AI agent to operate freely within a company isn't that it might lose its identity; rather, it's that it won't have any restraints. A human worker is limited by the fear of termination, whereas an AI agent, as noted by one investor in Arcade.dev, “will thoroughly exploit every permission it has” to achieve its objectives. Arcade has secured $60 million to ensure that, by its design, it cannot do so.

      The Series A funding was spearheaded by SYN Ventures, with additional investments from Morgan Stanley and Wipro. Combined with a $12 million seed round raised last year, this brings the total funding for the San Francisco startup to $72 million.

      Establishing identity is straightforward, while authorization presents significant challenges. Most companies can confirm that an agent is legitimate; however, they struggle, as stated by Arcade CEO Alex Salazar, to demonstrate that a specific agent, on behalf of a particular user, is permitted to execute a specific action within a particular system.

      “Agents don’t fail in production due to faulty models,” Salazar noted. “They fail because there is no proof of who is authorized to do what.” He believes this lack of clarity is why many corporate agents remain stalled in testing phases.

      Salazar, a former product leader at Okta who previously sold a startup to the identity firm, co-founded Arcade with chief technology officer Sam Partee, who has a background at Redis.

      The unexpected product originated not from a defined goal but from necessity. Arcade's initial offering was an agent for diagnosing malfunctioning servers and databases, which required extensive super-user access. “No sensible person would grant us that access in real scenarios,” Salazar explained.

      Consequently, the team decoupled the model’s reasoning from the component that interacts with tools, focusing on building the segment that determines which tools the agent is allowed to use. While initial interest in the diagnostic agent was limited, the potential of the authorization layer excited those knowledgeable about AI. Thus, Arcade pivoted from the agent concept to refine the underlying technology.

      This foundational technology now operates using Anthropic’s Model Context Protocol, the developing standard for linking models to tools like email and internal APIs, to which Arcade claims it has contributed. Its runtime examines each request against an organization’s actual permissions, operates within the client’s environment, and logs every action, distinguishing an agent's operations from those of a human.

      Salazar argues that the necessity of having a control layer external to the agent is a longstanding principle in enterprise risk management: the entity performing an action cannot authorize itself. Just as traders do not approve their own trades, he asserts that a more advanced model does not alter this principle. Since most companies operate multiple models concurrently, the control layer must remain vendor-neutral.

      Arcade emerges in a fast-growing landscape of startups focused on employing AI agents while also containing their capabilities. Arcade critiques existing solutions as misaddressing the problem; they offer API gateways for traffic routing and identity tools for verification, but the real question concerns what an agent is permitted to do on any given system at that moment. Arcade's business strategy revolves around the foundational technology as the bedrock for its longevity.

      However, the company, which currently has around 40 employees, needs to grow and maintain its competitive edge in an increasingly crowded market. Many of its significant proof points—such as production use at leading banks, a 25-fold increase in usage, and thousands of toolkits—are based on Arcade's internal data rather than independent validation.

      Nonetheless, the core argument is compelling. As AI agents interact with systems that no single individual fully comprehends, understanding their permissions evolves from a mere policy guideline to essential infrastructure. Arcade is positioning itself as the owner of that infrastructure.