Coursera's First Employee and a TensorFlow Co-Creator Raise $11 Million to Give Every AI Agent an ID Badge

Mintive, a startup building identity and access controls for enterprise AI agents, has raised $11 million in total, including a newly announced $7 million seed round led by Cowboy Ventures, the company said. Coatue, Operator Collective, Burst Capital, Hustle Fund and Maven Ventures participated, alongside angels including Okta co-founder Frederic Kerrest, WorkOS founder Michael Grinich, former Google chief scientist Jeff Dean and OpenAI co-founder Andrej Karpathy.

The founders explain the timing. Jiquan Ngiam, Mintive's chief executive, was Coursera's first employee, hired by Andrew Ng, and later worked on neural networks at Google Brain. Chief technology officer Vijay Vasudevan was one of the creators of TensorFlow at Google Brain. Both watched enterprises adopt agents faster than they could answer basic questions about them: which agent borrowed an employee's token, which one touched a production system, and who approved the action.

Mintive's answer is to treat an agent like a new hire. The platform issues each agent a scoped, revocable credential that is delegated from the person it serves, so an agent's permissions can differ from those of the employee who asked for the work. An MCP gateway approves which tool servers an agent may reach and can disable destructive tools by role; an LLM gateway routes model calls and caps spend per team; an agent monitor reconstructs the chain behind any action, including who initiated it, which server it hit and which model answered. A separate screen checks calls for prompt injection, leaked secrets and personal data before they reach company systems.

The company says the platform is model-agnostic, so Claude, ChatGPT, Copilot, Cursor and home-grown agents sit on the same identity and audit layer, and that it is SOC 2 Type II audited with EU and US data residency. Mintive launched its product in February and reports more than 50 paying customers, including Ro, Modern Treasury, the University of Michigan, Coursera, Apax and Stability AI, processing roughly 2.5 million agent calls a day and more than 200 million over seven months. It says Modern Treasury's bank-operations agent went live in under three weeks and saves technical account managers between 30 minutes and four hours per case, and that Michigan uses Mintive as an enterprise MCP gateway across about 150,000 community members.

Those figures are company-reported and have not been independently audited. The competitive field is also crowded: Amazon, Microsoft, OpenAI and Anthropic all sell agent management, point products come from LangChain, New Relic and Rippling, and more than 40 Y Combinator companies are working on adjacent pieces. Mintive's bet is that bundling the tool door, the model door, identity and audit into one product is easier to deploy than stitching four vendors together.

Its investor story is about restraint as much as technology. "AI is going to be able to power incredible things, but we need more infrastructure that's going to give us governance and visibility and control," Cowboy Ventures managing partner Aileen Lee said. She pointed to the size of the round as a reason she backed Ngiam, contrasting Mintive's $11 million total with agent infrastructure peers raising far more: Prime Intellect, for instance, raised $130 million at a $1 billion valuation. "He is a grown-up and so he was not a person chasing a vanity round," Lee said.

What this means: the agent governance market is forming around a simple test that most enterprises cannot yet pass, which is naming the actor behind a bad action in one click. Vendors that sell capability without that answer are selling a liability, and the buyers who will pay a premium are the ones whose auditors have already asked the question. Mintive is early, small and unproven at scale, but it has picked the right seam: the permission layer is where agent adoption either becomes infrastructure or becomes an incident.