What is proof of work for AI agents?

The phrase is borrowed, and that causes confusion. Ask most people what proof of work means and they will describe a mining rig burning electricity to win a lottery. That is a different idea with the same name. This article is about the ordinary English sense, where proof of work is the evidence that a piece of work was done.

That sense matters again because of who is doing the work. When a person moved a task to done, you judged the claim using everything you already knew about them. An agent gives you none of that. It will report success in the same confident sentence whether it wrote the migration or only described writing it. It will do that a hundred times a day.

Why a claim stopped being enough

A tracker is a record of claims. That was always true, and for twenty years it was fine. The claims came from people who had to sit in the same stand-up the next morning.

Two things break that. The first is volume. A team running agents produces more completed tasks per day than anyone will read carefully. Review turns into sampling, whether you planned it that way or not. The second is that the failure mode changed. A person who has not finished usually says so. A model that has not finished produces something that looks finished, because plausible-looking output is what it is best at.

So the question changes. It used to be "do I trust this worker". Now it is "what did this task produce, and can I look at it". That is proof of work.

What a proof of work contains

In Cognibl it is a set of files you can open:

• A CSV, filled in against the template for the process the project runs.
• The documents it references, which is usually where the substance is.
• A coverage report, where the claim involves test cases.

The CSV is what makes the proof checkable. A checker can confirm that the required columns are present and filled in before a person reads a single word. That check is cheap, and it filters out most of what would otherwise reach a human.

Harness, Graph and Loop each ship with a template. A project running a custom process defines its own. The gate enforces the template you defined, so a template that asks for little will refuse little. You decide how strict it is, and the report shows what was asked.

Every change makes a new version

Proof of work is versioned. A change creates the next version instead of editing the last one, so v1 stays readable next to v3. One version can hold several documents, because a real claim usually rests on more than one file.

The reason is practical. After something goes wrong, the useful question is what the evidence said when somebody approved the work. What it says today matters less. Only an append-only record can answer the first question.

Why the gate matters

Proof of work is only useful if something checks it. In Cognibl, a status change that completes a task will not apply without it. The refusal names what is missing, so nothing fails quietly.

Three layers do the checking, in order of cost:

1. Deterministic checkers, for anything a machine can decide outright.
2. An AI flow that reads the proof against the definition of done and writes up where they agree and where they do not.
3. A human sample, at a rate set by the agent's score, not by how anyone feels that day.

An agent can attach proof and request the status change. It cannot decide the outcome. There is also no code path that treats our own agents differently. A shortcut like that would be a bug.

What it is not

It is not a copy of your work. Your systems stay the system of record, so code stays in your Git and documents stay in your Drive. We store a reference and a content hash, captured at verification time. That ties every report to the exact bytes it saw, and there is no second copy anywhere to drift out of date.

It also does not replace writing down what you wanted. Proof of work is checked against a definition of done. If that definition is vague, the check will confirm that something happened. It will not tell you whether it was the right thing. That is the harder half of the problem, and it needs its own article.