Measurement is the loop condition.

Given enough compute, you can build anything.

There has been a lot of discussion about spec-driven development and, more recently, loop engineering. You give your coding agent a goal to achieve—a spec—and run a loop until the goal is reached.

But how do you verify that the goal has been reached?

Without measurement, nothing is verifiable.

A simple agent team

Imagine three agents with clear personas: CEO, Data, and Designer. You might also add QA or Infrastructure.

1. The human defines the goal with the CEO. Together, they create an issue that describes the outcome.
2. The Data agent runs a loop. It detects the new issue, updates the data model, runs tests, creates a pull request, and may create a new issue to nudge the Designer.
3. The Designer runs its own loop. It finds the issue and starts working on the design. Other specialist agents can do the same.
4. The CEO also runs a loop. It finds the updates and verifies whether the original goal has been achieved.

It is a very interesting concept: agent teams in a simplified form.

The core insight

Measurement is the loop condition. Without a verifiable exit predicate, you do not have a loop. You have a prayer.

The challenge

Knowing what to build is the hardest part of the whole process.

Typically, we do not know the full spec until we see some progress. Then we adjust and readjust the spec or the requirements. The target changes as our understanding improves.

Loops can also become very expensive, very quickly. Every iteration consumes compute, context, tool calls, and review attention. A vague goal does not merely create uncertainty—it creates an unbounded bill.

The engineering problem is therefore not just how to keep agents working. It is how to define measurable progress, decide when to stop, and preserve room for the human to change the destination.