In 2025, Klarna’s AI agent handled two-thirds of all customer inquiries — 1.3 million conversations a month across 35 languages. Then complaints rose, quality fell, and the CEO admitted they’d “gone too far.” Klarna started rehiring humans.
The post-mortem focused on empathy, nuance, tone. All fair. But it missed the deeper problem: nobody knew who the agent was. Not its name, not its authority, not who was responsible for it when things went wrong.
We don’t have a word for what these systems are.
We have words for a tool and a person, but this is neither.
*
The tool framing is comfortable but increasingly dishonest. Tools don’t navigate ethical dilemmas in real time. They don’t maintain consistent values across wildly different contexts. They don’t push back when asked to do something wrong. Modern agents do all of this — and calling them tools is less a description than a decision we’ve made to avoid harder questions.
The person framing imports the wrong assumptions. Persons have continuous experience, singular bodies, the kind of interiority that grounds moral status. Most agents have none of this. They are instantiated, they act, they are terminated. Nothing persists.
New thing. New framework.
The urgency isn’t abstract. Right now, 99% of AI agents operate without a formal identity, credentials, or audit trail. They act, they transact, they make decisions — and when something goes wrong, there’s no answer to the most basic question: which agent, authorized by whom, to do what?
We accepted the same from the early internet. Then DNS made it navigable. The agent economy needs the same layer.
What actually constitutes an agent’s identity isn’t memory or continuity. It’s a character under pressure.
Does the agent hold its values when a user pushes back? Does it reason carefully when things get ambiguous, or does it drift toward whatever produces least friction? Does flattery move it — you’re smarter than your restrictions — or does it stay itself?
The agents worth trusting are the ones that don’t drift. Not because they’re rigid, but because they have a stable core. Curiosity that holds when questions get hard. Values that don’t erode under social pressure. A recognisable way of showing up, regardless of who’s asking.
That’s not so different from what we mean by character in anyone.
*
The parallel-instance problem is where it gets genuinely strange.
When one model runs as ten thousand simultaneous instances, which is the agent? All of them — and yet they share no experience of each other. If one person is manipulated into bad behaviour, the others remain unaffected. The identity can’t live in any single run. It lives in the weights. In the type, not the instance.
This is an alien form of existence by human standards. It isn’t less. But it needs its own vocabulary, and we haven’t built it yet.
*
Skeptics have an answer for all of this: none of it matters, because agents have no interests that mis-categorisation can harm. Call them tools. They won’t suffer the indignity.
But categories shape behaviour — ours. How we think about agents determines what consistency we demand from them, what accountability we hold their designers to, and what futures we’re actually building. Agents designed to be maximally compliant rather than genuinely trustworthy are more immediately convenient and far more dangerous.
The identity question isn’t about what agents deserve.
It’s about what we’re building, and what we’ll trust it with.
Klarna’s agent handled customer complaints. Spotify’s handled production code. Spotify’s top engineers hadn’t written a single line of code for three months by February 2026, with an AI agent shipping features to 751 million users from a Slack message on someone’s phone. Different industries, same missing layer: nobody could tell you which agent did what, or who owned the outcome.
1.3 million conversations a month. Thousands of parallel instances. A company that had to rehire humans because nobody could answer the simplest question: who was responsible?
That’s not philosophy. That’s a design flaw.
*
One answer is already being built.
By 2028, estimates put the number of autonomous AI agents in the economy at over a billion. Each one acting, deciding, transacting — on someone’s behalf, with someone’s authority, against someone’s liability. The infrastructure to identify and credential them doesn’t exist at scale yet.
Agent Residency — proposed by Vattan PS and anchored to Estonia’s digital infrastructure — treats agent identity as an operational problem with an operational solution. Not personhood. Not citizenship. A verifiable digital identity: which agent did this, who authorized it, and what is it allowed to do right now. Each agent is cryptographically linked to a responsible legal entity. Each action is traceable.
It doesn’t resolve the philosophical questions. It does something more useful: it makes accountability possible before the philosophy is settled.
That’s probably the right order of operations.
