An agent task needs a visible outcome.

An agent can produce a convincing description of a completed task without changing anything. In a working environment, the important evidence is often elsewhere: a record was updated, a file contains the requested change, or the requested item appears in the correct place.

This distinction shapes how we think about agent data. A task needs an initial state, an instruction, the tools available to act on that state, and a completion condition that can be inspected. Without those pieces, a successful-looking sequence may be hard to distinguish from an incomplete one.

Useful environments also make room for ordinary friction. A tool can return an unexpected result. An instruction can leave a detail unresolved. The system may need to verify a previous action before taking another. These situations help define the decisions an agent is expected to make.

A demonstration should preserve enough context to explain those decisions. The action sequence alone is not always sufficient: what was visible, what had already been established, and which constraint applied can matter just as much as the next click or tool call.

Our interest is in learning material that keeps this connection intact. The instruction establishes the objective, the environment makes action possible, and the final state supplies evidence. Together, they provide a more useful basis for discussing whether a system completed the work.