AI agents need their own ID badges, not borrowed logins
AI agent identity is the idea that autonomous systems should operate with their own scoped, revocable credentials instead of masquerading under a human user’s account, giving each agent distinct permissions, audit trails, and shutdown controls that match its role in a production environment.
Vercel’s acquisition of Better Auth is a clear bet that this shift from borrowed logins to independent AI agent credentials will define the next phase of automation. Today, most agents open pull requests, ship deployments, or query internal tools while “wearing” the user’s badge, so every downstream service sees the person, not the agent. That is more than a nuisance; it is a structural security flaw. There is no clean way to limit what a given agent can do or to shut off only the misbehaving one without cutting off everyone else. In other words, we have human-grade identity in place while workloads are rapidly becoming machine-driven—and the mismatch is now impossible to ignore.
Agent Auth: turning AI agents into first-class, accountable actors
Better Auth, known for an open source TypeScript authentication framework with about 4.7 million weekly npm downloads, was already stretching beyond human login flows when Vercel came calling. Its emerging Agent Auth protocol aims to give AI agents identities of their own, with scoped, delegated, and revocable permissions separate from the person who deploys them. That is autonomous agent authentication by design, not a retrofit.
The opinionated move here is to stop treating agents as “features” on a user account and instead as independent actors with their own AI agent credentials and accountability. Vercel is folding this work into Vercel Connect and Eve, its products for connecting AI agents with external services, signaling that AI agent identity is not a side project but a core interaction layer. Importantly, Vercel says Better Auth remains open source under an MIT license, with the same contribution model and framework support, which keeps the protocol’s future from being locked behind a single SaaS wall.
Six million deployments a day: why identity suddenly matters
Vercel’s move would be easy to dismiss as a niche security feature if not for the scale of agentic workloads already in flight. Its infrastructure supports about 6 million deployments per day, with roughly half triggered by coding agents, and more than 1 trillion tokens flowing through its AI gateway daily. This is no longer a lab experiment; it is the production backbone for automated software changes.
Guillermo Rauch argues that the industry has shifted from prototyping AI agents to solving production problems, with coding agents and internal corporate agents emerging as dominant use cases. On that backdrop, letting agents impersonate humans is untenable. You cannot audit which subagent pushed a risky change if every action is tied to the same human identity. You cannot apply fine-grained limits or time-bounded scopes if the only knob you have is the person’s account. Agent infrastructure security has to evolve from role-based access for employees to capability-based access for machines that may spin up, run, and retire without a human ever logging in.
From experimental bots to an AI-to-AI interaction fabric
Better Auth’s founder Bereket Engida admits the agent identity problem is too large and fast-moving for a small team, which is why he chose to join Vercel’s broader infrastructure and distribution. That candid assessment matters: as autonomous agents move from experiments to production, identity is no longer a library concern; it is an ecosystem concern. Vercel Sandbox, for example, is already designed to restrict what data agents can access or export to prevent sensitive data from being swept into AI training.
Elsewhere in the ecosystem, Anthropic’s Claude Tag gives its model a distinct presence in Slack with its own connected accounts, instead of borrowing the identity of whoever invoked it. Infrastructure providers are quietly building a machine-to-machine interaction layer tailored to autonomous AI agent ecosystems, and Vercel is positioning itself as a central player in that stack as platforms and AI labs increasingly overlap in capabilities. The pattern is consistent: agents get names, keys, scopes, and logs. That is what AI agent identity looks like when it grows up.
The coming accountability layer for autonomous AI
The uncomfortable truth is that we shipped a generation of AI tools without a serious plan for accountability. When an agent acts on your behalf today, it usually runs under your identity and access, so every service it touches sees you, not the agent, and there is no clean way to isolate or disable one misaligned agent without collateral damage. That made sense when agents were toys. It does not when half of 6 million daily deployments are initiated by machines.
Giving agents their own scoped, revocable AI agent credentials is less about novelty and more about restoring basic security expectations: least privilege, clear audit trails, and targeted shutdowns. Agent Auth and similar efforts are early drafts of that accountability layer. If they succeed, we will stop asking whether autonomous agent authentication is necessary and start asking why any production agent is still running on a borrowed human badge.
