From Borrowed Logins to Agent Identity
AI agent authentication is the practice of giving autonomous software agents their own digital identity and scoped permissions so they can act in production systems without inheriting a human user’s full access, reducing the risk of privilege escalation and making their actions easier to audit and revoke when needed. That shift is the real story behind Vercel acquiring Better Auth, whose TypeScript authentication framework already sees around 4.7 million weekly npm downloads. Vercel is not buying login plumbing; it is buying a new security primitive: agent-specific identity. Today, most agents open pull requests, trigger deployments, or query internal systems using the same ID badge as the person who set them up. Every downstream service sees the human, not the agent, which means one misconfigured automation can quietly wield full user power. Vercel’s move is a clear bet that this “borrowed login” model is untenable at scale and that enterprises will demand a distinct agent identity protocol as aggressively as they now demand multi-model routing.
What Agent Auth Changes About Enterprise Agent Security
Better Auth’s emerging Agent Auth protocol gives AI agents identities of their own, with scoped, delegated and revocable permissions separate from those of the person deploying them. In plain terms, each agent stops being a shadow extension of a user account and becomes a first-class actor with its own access envelope. That is not a minor tweak; it rewrites how enterprise agent security should work. “When an agent acts on your behalf, it runs under your identity and access, so every service it touches sees you, not the agent,” Guillermo Rauch points out, highlighting the current inability to limit one agent or shut it down without cutting off the rest. Scoped permissions agents flip this logic: you can grant task-specific rights, cap what data an agent can touch, and revoke its credentials without tearing down an entire application’s access graph. The result is fewer opportunities for unauthorized actions and a more honest audit trail of what the software, not the human, did.
Task-Specific Access Beats Blanket Trust
Enterprises do not run a handful of AI helpers; they run fleets. Coinbase already operates roughly 1,200 full-time AI agents by normalizing compute hours to a standard workweek, and its CEO argues that human developers have no business manually picking models at that scale. When hundreds or thousands of agents commit code, query ledgers, or poke at CRM systems, blanket application access is reckless. Agent Auth’s scoped credentials let teams grant permissions that match a concrete job: this agent opens pull requests only in one repository, that agent reads deployment logs but cannot change infrastructure, another can update a narrow set of business records. Because permissions are delegated and revocable, security teams gain a clean kill switch for a single misbehaving agent instead of yanking a user’s entire account. More importantly, they gain visibility: each identity maps to a specific agent, making it far easier to audit behavior than sifting through logs where everything appears under a human’s name.
Identity as the Missing Control Plane
The industry has spent the last year obsessing over model choice and inference costs while quietly ignoring who, exactly, the agents appear to be. Vercel now routes more than a trillion tokens a day across millions of deployments, and is moving away from single-lab partnerships because models have become interchangeable components in a larger pipeline. Coinbase cut internal AI spend nearly in half by defaulting engineers to cheaper open-weight models, relying on an internal gateway to route tasks and push cache hit rates from 5% to 60%. These strategies turn gateways into cost and performance control planes, but they still sit on fragile identity foundations if every agent acts as a user proxy. Vercel’s acquisition is a statement that governance must catch up: AI agent authentication and an agent identity protocol need to be baked into the same infrastructure that optimizes tokens. Without scoped identity, routing logic saves money while leaving enterprises exposed to opaque, over-privileged automation.
What Comes Next: Agents Treated as First-Class Citizens
The acquisition is not a side project; Better Auth creator Bereket Engida and the core team are joining Vercel to keep pushing both the framework and agent identity work. That effort will now feed into Vercel Connect and Eve, products aimed at wiring AI agents into external services. In other words, the company is preparing for a world where agents are standard participants in production traffic, not experimental add-ons. Engida is blunt about why this needed a bigger platform: “We’re a small team; the agent identity problem is big and moving fast,” he notes, arguing that Vercel’s infrastructure and community are required to bring Agent Auth to developers at meaningful scale. The conclusion for enterprises is equally blunt. If agents will open pull requests, review code, and touch business systems at scale, then giving them first-class identities with scoped, revocable permissions is no longer optional. It is the difference between safe automation and a swarm of invisible superusers.
