Every major AI lab is shipping agents right now - Anthropic has Claude’s computer use, OpenAI has Operator, Google has Project Mariner - and none of them agree on what an agent actually is. That sounds like a pedantic concern. It isn’t.
The Definition Problem Is a Product Problem
When ‘agent’ can mean anything from ‘a chatbot that fills out a web form’ to ‘an autonomous process that runs for hours without supervision,’ the word stops carrying information. Users can’t calibrate expectations. Enterprises can’t assess risk. And developers building on top of these platforms are making architectural decisions based on capabilities that may shift quarterly without warning.
This would be fine if the stakes were low. They’re not. Companies are already deploying AI agents inside internal tooling, customer-facing workflows, and, in some documented cases, against production databases. The gap between what a demo shows and what a deployed agent actually does in an edge case is the gap where things go wrong.
What Labs Are Actually Shipping
The current generation of agents is, mostly, wrapper logic around existing models - sequences of tool calls, memory retrieval, and browser or API access stitched together with prompting. That’s not a dismissal; the engineering to make it reliable at scale is genuinely hard. But it means the ‘intelligence’ in the agent is still the base model’s intelligence. When the model hallucinates a function call, or misreads a page state, or loops on a subtask, the agent fails in ways that are much harder to catch than a bad chatbot response.
Anthropomorphising agents - calling them ‘autonomous,’ framing them as digital workers - makes this harder to debug, not easier. Users who believe the agent understands intent will blame themselves when it doesn’t, rather than the product.
The Accountability Gap
Here’s the part nobody has solved: when an agent takes a wrong action - sends an email it shouldn’t, deletes a record, makes a purchase - the question of who is responsible is genuinely open. Terms of service tend to place liability on the user. But users were told the product was ready.
There’s no clean answer here yet, and the labs know it. What’s less clear is whether the race to ship is creating pressure to avoid asking the question too loudly - or whether the answer will only start taking shape once something goes visibly, expensively wrong at scale.