{"slug":"what-is-an-ai-agent","title":"What is an AI agent?","summary":"The word changed meaning in 2025, and it changed toward what these things can do. The question worth asking is who is answerable when one of them is wrong.","industry":{"slug":"groundwork","label":"Groundwork"},"url":"https://www.animyst.ai/field-notes/what-is-an-ai-agent","published":"2026-09-17","updated":"2026-09-17","author":"Abhi Chatterjee","blocks":[{"type":"paragraph","text":"Some time in 2025 the word settled. It moved from the academic framing (systems that perceive, reason and act) to the one the industry now uses: [large language models capable of using software tools and taking autonomous action](#s:definition-shift)."},{"type":"paragraph","text":"Both of those describe what the thing is *able* to do. Neither says how often it does it correctly, and neither says what happens when it doesn't."},{"type":"heading","text":"What the word actually carries"},{"type":"paragraph","text":"Nobody outside this trade says \"agent\" to name an architecture. They reach for the older sense of the word: something that acts on your behalf."},{"type":"paragraph","text":"On its own that narrows very little. A script acts on your behalf. So does a scheduled report, and so does the person who opens the post. The word is carrying more than that, and it is worth being precise about what."},{"type":"paragraph","text":"Two things. The first is judgement. You do not hand an agent a script, you hand them an outcome and let them work out the steps. A travel agent decides things for you, and that is the entire point of using one."},{"type":"paragraph","text":"The second is accountability. An agent is answerable to somebody. They act on your behalf, which means that when they get it wrong, it lands on you."},{"type":"paragraph","text":"The first is the part people are excited about. The second is the part they are nervous about, and it is the one nobody in the category is answering. So the question worth asking was never what an agent is. It is this: if this thing acts on my behalf, who is answerable when it is wrong?"},{"type":"heading","text":"Capability is not reliability"},{"type":"paragraph","text":"Here is the part the category does not lead with. A group at Princeton put it plainly this year: capability and reliability are separate properties, and improving one does not improve the other."},{"type":"quote","text":"A highly capable system can be unreliable, and a less capable system can be highly reliable within its envelope. Improving capability does not automatically improve reliability.","attribution":"Towards a Science of AI Agent Reliability, ICML 2026"},{"type":"paragraph","text":"They measured it. Across fifteen models and two benchmarks, [recent capability gains produced only small improvements in reliability](#s:reliability-science). Reliability barely moved at all over two years of releases."},{"type":"figure","value":"24 months","caption":"Of model releases, over which reliability showed minimal improvement. Every frontier provider clusters in the same place, so this is an industry-wide plateau rather than one vendor's problem."},{"type":"paragraph","text":"So the thing being advertised and the thing you need are not the same thing, and getting more of the first does not get you the second. Buy on capability and you are buying the wrong property."},{"type":"heading","text":"So what are you buying"},{"type":"paragraph","text":"The outcome. That is the only real signal. Everything else is noise: the capability, the model, the word itself."},{"type":"paragraph","text":"Easy to say. Harder to answer, because if the part in the middle is not getting more dependable, something else has to be doing the work. Three things, and none of them is the model:"},{"type":"flow","steps":[{"label":"Guardrails","detail":"It cannot reach the things you could not come back from."},{"label":"Narrow scope","detail":"One job, drawn tightly enough that the blast radius is survivable."},{"label":"A human on anything irreversible","detail":"A person at every one-way door."}]},{"type":"paragraph","text":"Look at what those three have in common. Not one of them makes the model more accurate. Every one of them makes *being wrong survivable*."},{"type":"heading","text":"Not yet, and what would change it"},{"type":"paragraph","text":"Which brings it back to the boundary. What we would not hand over is a decision that is load-bearing to a process or a business. Not yet."},{"type":"paragraph","text":"That phrase is doing real work, and it is worth saying what it means. It does not mean waiting for a better model: the research above suggests that wait could be long and might not help. It means not until the controls around that decision are good enough that being wrong is recoverable."},{"type":"paragraph","text":"Which is what load-bearing was always describing. A decision is load-bearing when getting it wrong brings something down. That is the same sentence as saying you cannot come back from it."},{"type":"paragraph","text":"So the line moves. It moves as you build, not as the model improves. And it answers the question the word raised at the start: you are still the one answerable. The work is making sure that being wrong is something you can come back from."},{"type":"callout","label":"The short version","text":"An agent acts on your behalf, using its own judgement, and you stay answerable for it. Whether it is worth having comes down to what it does every time, not what it can do once."},{"type":"qa","items":[{"question":"What is an AI agent?","answer":"In the industry's current sense, a large language model that can use software tools and take action on its own. In the older and more useful sense, something that acts on your behalf, using its own judgement, for which you remain answerable."},{"question":"Does a more capable model make a more reliable agent?","answer":"No. They are independent properties. Across fifteen models measured on two benchmarks, capability has climbed while reliability has stayed roughly flat for two years, and every frontier provider sits on the same plateau."},{"question":"How do you get a dependable outcome out of an unreliable model?","answer":"Not from the model. From guardrails, from scope narrow enough that the job can be done right, and from putting a human on anything irreversible. None of those makes the model more accurate. All of them make being wrong survivable."},{"question":"What should you not hand to an AI agent?","answer":"Decisions that are load-bearing to a process or a business, for now. A decision is load-bearing when getting it wrong brings something down, which is another way of saying you cannot come back from it. The boundary moves as the controls around a decision get good enough that being wrong becomes recoverable."}]}],"sources":[{"id":"definition-shift","title":"AI agents arrived in 2025 — here's what happened and the challenges ahead in 2026","publisher":"Thomas Şerban von Davier, The Conversation","url":"https://theconversation.com/ai-agents-arrived-in-2025-heres-what-happened-and-the-challenges-ahead-in-2026-272325","retrieved":"2026-09-17"},{"id":"reliability-science","title":"Towards a Science of AI Agent Reliability","publisher":"Rabanser, Kapoor, Kirgis, Liu, Utpala and Narayanan, ICML 2026","url":"https://arxiv.org/abs/2602.16666","retrieved":"2026-09-17"}],"disclosure":"These are Abhi's positions, taken down in an interview each week and checked against the sources listed with the piece.","questions":[{"question":"What is an AI agent?","answer":"In the industry's current sense, a large language model that can use software tools and take action on its own. In the older and more useful sense, something that acts on your behalf, using its own judgement, for which you remain answerable."},{"question":"Does a more capable model make a more reliable agent?","answer":"No. They are independent properties. Across fifteen models measured on two benchmarks, capability has climbed while reliability has stayed roughly flat for two years, and every frontier provider sits on the same plateau."},{"question":"How do you get a dependable outcome out of an unreliable model?","answer":"Not from the model. From guardrails, from scope narrow enough that the job can be done right, and from putting a human on anything irreversible. None of those makes the model more accurate. All of them make being wrong survivable."},{"question":"What should you not hand to an AI agent?","answer":"Decisions that are load-bearing to a process or a business, for now. A decision is load-bearing when getting it wrong brings something down, which is another way of saying you cannot come back from it. The boundary moves as the controls around a decision get good enough that being wrong becomes recoverable."}],"contact":"https://www.animyst.ai/talk"}