AI agents are becoming first‑class identity holders
AI agent identity is the idea that autonomous AI agents should authenticate and be authorized as distinct digital actors, with their own scoped, revocable permissions and machine-to-machine credentials instead of reusing a human user’s account or session tokens in production systems.
That may sound academic, but it is now an infrastructure problem at scale. Vercel supports 6 million deployments every day, and roughly half are triggered by coding agents rather than humans clicking deploy. When software is deploying software, pretending that all actions belong to a human user is no longer harmless bookkeeping; it is a security bug. Today, most autonomous AI agents open pull requests, query internal systems, and update apps while “wearing” their creator’s identity. That makes it impossible to see which agent did what, limit individual agents, or revoke access without collateral damage. The acquisition of Better Auth to build Agent Auth is a bet that identity for agents will be as central as identity for people.
Vercel’s Agent Auth: giving agents scoped, revocable IDs
Vercel is acquiring Better Auth, whose open source TypeScript authentication framework sees about 4.7 million weekly npm downloads, and bringing its creator Bereket Engida and core team in-house. Better Auth began as a way to add login, sessions, user management, and permissions to web apps across frameworks, but the team had already moved toward agent-specific identity before the deal.
Their answer is Agent Auth, an open agent authentication protocol that gives AI agents their own identities with scoped, delegated and revocable permissions, separate from the human who launches them. This is agent authorization designed around autonomous software, not people. Instead of every request being stamped with a user, each agent becomes a distinct principal whose rights can be minimized, audited, and killed off independently. Guillermo Rauch has been blunt: current identity systems were built for people, so when an agent acts, every service still thinks it is the human, leaving no clean way to limit or shut down one agent or subagent without cutting off the rest.
From borrowed logins to real machine-to-machine infrastructure
The uncomfortable reality is that most autonomous AI agents today operate under borrowed logins. Every tool an agent touches sees the person, not the agent, which makes fine-grained control almost impossible. In small prototypes this feels acceptable; in production, with half of 6 million daily deployments triggered by coding agents, it is a risky blind spot. As workflows grow more autonomous, developers need a machine-to-machine infrastructure layer that understands agents, not just humans.
Agent identity protocols are a concrete way to secure delegation. If an internal automation agent can deploy services, query CRMs, and export reports, each of those capabilities should be expressed as narrow, revocable permissions attached to the agent’s own credentials, not a cloned human admin token. That makes it possible to see what each agent did, to restrict its resource access to the minimum, and to shut down only the misbehaving one. This is the same principle that improved cloud security when teams stopped sharing root keys and started giving services their own identities; agents deserve the same treatment.
Why now: agents have escaped the lab
The timing is not an accident. According to Guillermo Rauch, the industry has moved from prototyping AI agents to solving production problems, with coding agents and internal corporate agents becoming the dominant use cases. Better Auth’s Engida is blunt about the scope: “We’re a small team; the agent identity problem is big and moving fast,” he wrote in his blog post explaining why the company joined Vercel.
Vercel is already building around this reality. Its Eve framework lets companies define agent instructions and skills in natural language, while Vercel Sandbox restricts what data agents can access or export, aiming to reduce the risk that sensitive data is pulled into model training or exposed where it should not be. The plan is for the Agent Auth work to feed into Eve and Vercel Connect, the product that wires agents to external services. In other words, identity is being wired directly into how agents are defined, what they can see, and how they call other systems. This is not a bolt-on feature – it is the control plane for autonomous behavior.
What this means for developers building with autonomous AI agents
For developers, the message is clear: treating agents as invisible extensions of users will not survive contact with real workloads. Better Auth will remain MIT-licensed and open, with the same community governance and framework support even under Vercel’s ownership, which means the tools for modern AI agent identity are likely to spread fast across ecosystems. At the same time, providers are building systems like Eve, Sandbox, and Connect that assume agent-to-agent interactions and machine autonomy as first-class patterns, not edge cases.
The practical impact is straightforward. If you are shipping AI agents that touch code, data, or business systems, you will be expected to show which agent did what, limit each agent’s blast radius with scoped permissions, and revoke their access without collateral damage. Better Auth’s work on Agent Auth, backed by Vercel’s scale, is likely to set early norms for AI agent identity, agent authentication protocol design, and agent authorization semantics. Developers who design for this model now will be better prepared as infrastructure platforms and AI labs continue to overlap and race to support more autonomous software.
