Your identity stack was built for two kinds of actor. Agents are a third.

Why AI Agents Need Their Own Identity Stack (And What That Means for Security)

For decades, the digital world has recognized exactly two kinds of actors. Humans log in with passwords, biometrics, or single sign-on. Services and applications authenticate with API keys, OAuth tokens, and service accounts. These two identity models have shaped everything from enterprise security to consumer apps. But a third kind of actor is emerging — one that doesn't fit neatly into either box. AI agents are arriving, and they bring a fundamental challenge: your identity stack was not built for them.

The moment an AI agent starts acting on your behalf — booking travel, managing your calendar, negotiating with vendors, writing code, or even making financial transactions — it needs a verifiable identity. It needs permissions. It needs audit trails. And it needs a way to prove to every system it touches that it has the authority to act. Today's identity infrastructure is not designed for that. Let's explore what this shift means and how the identity stack must evolve.

The Two-Actor Identity Model That Defined the Internet

To understand the problem, we need to appreciate how identity worked until now. The digital identity stack was built around two primary actor types, each with its own patterns, protocols, and assumptions.

Human actors are what we all experience daily. A human logs into a system using a username and password, a magic link, a biometric scan, or a social login. Once authenticated, the system knows who that person is — and more importantly, what they are allowed to do. Humans have limited attention spans, operate in real time, and typically act on one session at a time. They can be prompted for consent, challenged with multi-factor authentication, and held accountable because they are individual people with reputations and legal standing.

Service actors are the other category. These are applications, microservices, APIs, and backend systems that talk to each other. They authenticate using client credentials, API keys, jwt tokens, and OAuth flows. Service actors are automated, follow strict rules, and have clearly scoped permissions. They don't need to be prompted for consent — they are configured with permissions upfront by developers or administrators. Their behavior is (in theory) deterministic and predictable.

These two models have served well for a long time. Humans get the flexibility they need, and services get the automation and security they require. The identity stack — including protocols like OAuth 2.0, OpenID Connect, SAML, and SCIM — evolved to serve these two actor types. But AI agents break this neat binary.

Enter the Third Actor: The AI Agent

AI agents are something new. They are automated, like services. But they are autonomous, like humans. They can make decisions, choose between multiple actions, and act on behalf of a person or organization without being explicitly instructed in every moment. They learn from context, operate across multiple systems, and can run continuously without human supervision.

This combination creates an identity crisis. An agent needs permissions that are flexible enough to handle unexpected situations, but constrained enough to prevent abuse. It needs to prove its identity across services — not just once, but continuously. And when something goes wrong, someone needs to be accountable. But who? The person who set up the agent? The developer who wrote its code? The organization that deployed it?

Consider a practical example. A personal AI agent is tasked with managing your financial life. It monitors your bank accounts, pays bills, moves money between savings and checking, negotiates with credit card companies, and files insurance claims. Today, each of those services has identity protocols designed for either a human (you logging in) or a service (your bank's API using OAuth). An agent doesn't fit either pattern. It isn't you — it can't present your biometrics or answer your security questions. And it isn't a pre-configured service — it needs to adapt its actions based on your changing needs.

What the New Identity Stack Must Do

Building identity for AI agents is not a simple patch to existing protocols. It is a foundational shift. Here are the core capabilities that an agent-ready identity stack must provide.

Delegated Authority with Granularity

Agents must be able to prove that a human or organization has authorized them to take specific actions. But that authorization needs to be far more granular than "yes" or "no." It needs to include scope (which accounts, which transaction types, which time windows), constraints (maximum amounts, frequency limits, allowed recipients), and conditions (only during business hours, only if the human is not currently overriding it). Think of it as a power of attorney — but for digital actions, and revocable in real time.

Continuous Authentication and Authorization

For humans, authentication happens at login. For agents, it must be continuous. An agent may run for days or weeks, interacting with dozens of services. Its authority must be checked at every step — not just once. This means identity tokens with very short lifetimes, context-aware permission checks, and the ability to revoke access instantly when unusual behavior is detected. The stack must support real-time authorization decisions, not just pre-configured roles.

Auditability and Accountability

Every action taken by an agent must be logged in a way that ties back to its identity and to the human or organization that authorized it. This goes beyond standard audit logs. It requires chain-of-authority records — who delegated what to whom, and when. When an agent makes a mistake, there must be a clear trail showing which delegation path led to that action. This is essential for compliance, legal liability, and trust.

Human-in-the-Loop Patterns

Not all agent actions should be automatic. For high-stakes decisions — large financial transfers, contract signings, data deletion — the identity stack must support human approval workflows. But these must be designed for agent contexts. A human should not need to log in to a separate system each time. Instead, the agent should be able to request approval through the same identity channel, with the human approving via a simple push notification or voice confirmation, and the agent receiving a cryptographically signed authorization token.

Agent-to-Agent Authentication

Agents will need to negotiate with each other. Your personal shopping agent will interact with a retailer's customer service agent. They need to mutually authenticate, verify each other's authority, and establish trust — without human intervention. This is a new protocol space entirely. It requires agent identity certificates, reputation systems, and standardized handshake protocols that work across organizations and trust boundaries.

Implications for Businesses

For organizations, the arrival of agent identity is both a risk and an opportunity. The risk is obvious: agents that lack proper identity controls will become vectors for fraud, data breaches, and compliance violations. An agent that can impersonate a user or a service without being properly accountable is a security nightmare.

The opportunity is equally significant. Businesses that build agent-ready identity infrastructure will be able to offer new services that were previously impossible. Imagine an insurance company that offers a claims agent — an AI that can access policy details, chat with customers, authorize payments, and coordinate with repair services, all within a secure identity framework. Or a bank that offers a financial wellness agent that can negotiate with creditors, optimize spending, and manage accounts across multiple institutions.

But to get there, businesses need to start thinking about identity differently. The identity team, the security team, and the AI team need to work together. Existing identity vendors will need to add agent-specific flows. New standards will need to emerge. And organizations that wait for perfect protocols will fall behind those that start experimenting now.

Practical Steps for Business Leaders

Audit your current identity stack for agent readiness. Where would an agent acting on behalf of a user or service hit a wall? Which permissions cannot be delegated granularly? Where are audit trails insufficient?
Start piloting agent identity patterns in low-risk, internal use cases. Give a customer support agent limited access to knowledge bases and ticketing systems using short-lived, scoped tokens. Learn what breaks.
Demand agent support from identity vendors. Ask your IAM provider: When will you support delegated authority for AI agents? When will you support continuous authentication for non-human actors? If they don't have a roadmap, consider alternatives.
Develop internal standards for agent accountability. Every agent deployed in production should have a unique identity, a defined scope of authority, a human owner, and an audit trail that is as rigorous as any human employee's.
Engage with standards bodies. Groups like the OpenID Foundation and the FIDO Alliance are beginning to explore agent identity. Your organization's voice — and use cases — can help shape the protocols that will govern this space.

Implications for Society

The agent identity problem is not just technical. It has broad societal implications. If agents are not properly authenticated and accountable, we risk a world where scams are automated at scale, where agents impersonate people to commit fraud, and where nobody can be held responsible for the actions of autonomous systems.

On the other hand, if we get agent identity right, we open the door to a new era of digital autonomy. People can trust agents to act on their behalf across the entire internet. Small businesses can deploy agents to negotiate with suppliers, manage logistics, and handle customer service — competing with larger companies that have entire teams for these tasks. Elderly or disabled individuals can have agents that manage their daily digital lives with appropriate safeguards.

Regulation will play a role here. Governments are already beginning to ask questions about AI accountability. The identity stack is the technical foundation for that accountability. A regulatory framework that requires agents to have verifiable identities, scoped authorities, and auditable actions is likely inevitable. Forward-thinking organizations should treat compliance not as a burden, but as a design constraint that builds trust.

What This Means for the Future of AI and How It Will Be Used

The rise of agents as a third actor type is one of the most consequential shifts in computing since the transition from mainframes to personal computers, or from desktop to mobile. Identity is the critical layer that will determine whether agents become a trusted utility or a security catastrophe.

In the near future, every person will have multiple agents acting on their behalf — shopping agents, scheduling agents, financial agents, health agents, creative agents. Each of those agents will need an identity that is distinct from the person's own identity, yet verifiably linked to it. The identity stack must support this without requiring the person to constantly approve every action, while still maintaining control and accountability.

For businesses, the winners in the agent era will be those that build identity infrastructure that is flexible, secure, and human-centered. The identity stack is not just a backend concern — it is the interface between humans and their autonomous digital representatives. Companies that make agent identity seamless and trustworthy will earn the most valuable currency in the AI age: user confidence.

For society, the path forward requires open standards, broad collaboration, and a willingness to treat agent identity as a first-class problem — not an afterthought. The two-actor model served us well, but it is no longer sufficient. The third actor is here. It is time to rebuild the identity stack for the age of agents.

TLDR: The traditional identity stack was built for two kinds of actors — humans and services — but AI agents represent a third kind that does not fit either model. Agents need delegated, granular, and continuously verifiable authority along with robust audit trails and human-in-the-loop approval patterns. Organizations that begin adapting their identity infrastructure now — by auditing current capabilities, piloting agent-specific patterns, and engaging with emerging standards — will be better positioned to trust and scale agent-based systems. Getting agent identity right is not just a technical challenge; it is the foundation for the entire future of autonomous AI.