When coding gets cheaper, requirements become more valuable

AI coding agents are making implementation dramatically faster. But faster implementation doesn't necessarily mean faster delivery of the right thing.

A vague requirement can now become working software surprisingly quickly.

And that creates an interesting shift.

When writing code was expensive, ambiguity was often discovered gradually through conversations, implementation, testing and review. Agents remove much of that friction. They can produce a plausible solution before we've properly decided what the solution is supposed to achieve.

Markus Eisele argues that this changes the economics of specification. Vague prompts may look cheap upfront, but the cost reappears later as correction loops, human review and rework. Clear acceptance criteria move more of that verification upfront and can give agents an explicit target against which their output can be checked.

This doesn't mean returning to huge requirements documents.

The useful specification might be relatively small: the business intent, important constraints, examples, non-goals, edge cases and observable acceptance criteria. For well-understood behaviour, some of those criteria can even become executable tests.

There's another interesting possibility: agents themselves can help create and challenge the specification. One agent can draft requirements, another can look for ambiguity, contradictions and missing cases, while humans remain responsible for validating whether the specification actually represents the intended business outcome.

As implementation becomes cheaper, the bottleneck doesn't disappear.

It moves upstream.

Practical takeaway: don't only invest in making agents better at building software. Invest in making the problem, expected outcome and evidence of success clear enough that agents can build the right thing — and we can verify that they did.