Matt Pocock asked a good question on X today.
He said he needs a word for the type of task that agents are really good at. The kind where the agent has 100% of the data, it can run experiments, verify the result on its own, and hill climb towards its destination. On those tasks, he said, his trust is maxed out and the results are incredible. He just needed a word for them.
My answer was one word.
Agentable.
Here is the post where I put it, on October 27, 2026:
Agentable
Rami (@KingRomstar) October 27, 2026
This post is me staking out what the word means, so it has a definition attached to it and not just a vibe.
The definition
Agentable (adjective): a task is agentable if an agent has all the data it needs, can run experiments cheaply, can verify the result without asking a human, and has a clear score it can hill climb.
That is it. Four conditions. If a task has all four, hand it over and go do something else. If it is missing one, you are not delegating, you are babysitting.
The four conditions
1. The agent has the data.
Not most of the data. All of it, or at least all of it in places the agent can actually reach. The moment the real answer lives in a Slack thread the agent cannot see, or in your head, or in a PDF on someone's laptop, the task stops being agentable. The agent will not tell you a piece is missing. It will fill the gap with something plausible and keep going.
2. It can try things.
Experiments have to be cheap and reversible. Run the test suite. Try the query. Build the prototype, measure it, throw it away, build another one. Agents get good results by iterating, not by being smart in one shot. If every attempt costs real money or breaks something you cannot undo, iteration dies and so does the whole advantage.
3. It can check its own work.
This is the one people skip, and it is the most important one. There has to be a way to know the answer is right that does not involve a human reading it and nodding. Tests pass. The numbers reconcile. The output matches the source. Without independent verification you do not get an agent, you get a very confident intern who never says "I don't know."
4. It has a destination.
A metric. Acceptance criteria. A definition of done that fits in a sentence. "Improve conversion" is not a destination. "Get the failing tests from 40 to 0 without changing any test" is. Hill climbing only works if there is a hill and the agent can tell up from down.
What is not agentable
Plenty of work fails this test, and it is worth naming, because people keep throwing these tasks at agents and then blaming the model.
Tasks where the goal is taste. "Make it look nice." "Write it in my voice." There is no score, so there is no hill to climb.
Tasks where the data is people. Negotiating, reading a room, knowing which stakeholder is actually the decision maker. The data exists, but it is not written down anywhere an agent can get it.
Tasks that are one shot and irreversible. Sending the email is cheap. Unsending it is impossible. The experiment condition fails.
Tasks where verification is the whole job. If checking the work takes as long as doing the work, you saved nothing.
None of these are agent failures. They are category errors. The task was never agentable in the first place.
The useful part: agentability is a dial, not a switch
This is the part that actually matters in practice.
Most tasks are not agentable as they land on your desk, but they can be made agentable. You pull the relevant data into one place the agent can reach. You write the acceptance test before you hand the task over. You split the one irreversible step out and keep that step for yourself. You replace "make it better" with a number.
So the question stops being "can an agent do this?" That question has a useless yes or no answer. The better question is "what would make this agentable?" That question has a todo list attached to it, and most of the time the todo list is short.
The noun form, by the way, is agentability, and the opposite is unagentable. A task with high agentability is one you can hand over on Friday and review on Monday. A task with low agentability is one where you will spend more time supervising than you would have spent doing it yourself.
Use the word
Words only stick if people use them, so use this one. Next time a task lands in front of you and it has the data, the cheap experiments, the self check, and the score, you do not need a paragraph to describe it.
It is agentable. Hand it over.
Defined by Rami Rashid. First posted on X on October 27, 2026: https://x.com/KingRomstar/status/2115096984420028788