A task an agent can finish alone: every input is in hand, every attempt can be checked, and every attempt scores higher than the last. Hand it over and walk away.
Give the agent the data, a way to act, and a test it can run on itself. The loop closes without you, so the trust is not a leap of faith. It is a measurement.
One email when the first task types open. Nothing else.
Four conditions. All four, or it is not agent-complete.
A task earns the name when the loop can close without a human in it. Miss one condition and you have an open task, which means you are the verifier, and you will be babysitting.
Complete inputs
Every fact the task needs is already in hand, in a form a machine can read. No hidden spreadsheet, no missing registry, no "ask Steve".
It can act
The agent can run the experiment: execute code, edit the file, call the API, take the measurement. Reading the data is not doing the task.
It can check itself
A test decides pass or fail with nobody looking. A number, a diff, a green suite. Not a vibe, not a screenshot someone eyeballs.
It hill climbs
Each attempt produces a score and the score moves up. The loop stops when the score stops mattering, or when it stops climbing.
Trust is a measurement, not a promise.
On an open task you are the verifier, so every attempt costs you attention. On an agent-complete task the verifier is code, so the agent can run the loop a thousand times before you look at it once.
That is why the trust feels total. The agent is not guessing at what good looks like, it is scoring against it, and the line on the right only goes up.
Where it has 100% of the data, it can run experiments, verify independently, and hill-climb towards its destination. On these tasks my trust level is totally maxed out.
Matt Pocock, asking for the word. 27 Oct 2026.
Is your task agent-complete?
Four switches. Flip the ones that are actually true today, not the ones that will be true once someone builds something.
A name is only useful if it changes the next decision.
Three things a shared word unlocks. All of them start with the same question, asked before a model or a framework is chosen.
The first question
"Is this agent-complete?" gets asked while scoping the work. If the answer is no, you budget for review. If the answer is yes, you budget for the verifier and then get out of the way.
A registry of task types
Named task classes that close on their own, each with the test that closes it and the score that shows it climbing. Sorted by how much trust the evidence actually earns.
Agents for one class
An agent tuned for a single agent-complete class beats a general agent asked to do everything. The definition is the spec.
Get on the list.
The definition, the four-condition checklist, and the first task types as they open. One email, then silence until there is something real.
No spam. Unsubscribe in one click.