Spec-first AI development is quietly rebuilding the one thing engineers spent a decade escaping: waterfall.
The agent doing the typing won't save you from it — if anything, it makes the trap worse.
That's why I'm sceptical of approaches like Spec Kit becoming the default way teams work with coding agents. Not because specs are bad. Specs are useful. Intent matters, acceptance criteria matter, boundaries matter. But the moment the answer becomes "write a bigger upfront spec so the agent can execute," you've recreated the exact workflow most good engineers walked away from.
And they walked away for a reason. Spec-first asks you to answer too much before you know enough. You're expected to define the full shape of the work before the code has been inspected, before the edge cases surface, before the awkward legacy behaviour is found, before the real rabbit holes show up. That's why agile, prototyping, small slices, and fast feedback won in the first place. Software is full of unknowns, and you cannot spec an unknown.
With agents the risk gets sharper, not softer. Too thin a spec and the agent fills the gaps. Too heavy and it confidently executes assumptions that were already wrong. One stale instruction, one over-confident leap, and the work spirals — faster than a human ever would.
Here's the part that gets missed: the iterative approach doesn't just protect quality at the cost of speed. I'd argue it delivers closer to what was actually wanted and gets there quicker — because you're correcting course early instead of discovering at the end that the spec was wrong. And what comes out the other side is easier to maintain, change, and scale, because it was shaped by the real system rather than an upfront guess.
So I'm not trading away engineering judgement to buy a little speed. The speed should adapt to the proven workflow, not the other way round.
You don't hand a senior developer a 40-page implementation script. You give them the outcome, the guardrails, the standards, the risks — then you expect them to inspect the system, use judgement, and bring back evidence. Agents are no different.
The future of AI delivery isn't bigger prompts or heavier process. It's just enough structure to stay safe, enough freedom to do useful work, and enough verification to know the work can be trusted.
Not vibe coding. Not AI waterfall. Disciplined, governed, iterative delivery.
The agent doing the typing won't save you from it — if anything, it makes the trap worse.
That's why I'm sceptical of approaches like Spec Kit becoming the default way teams work with coding agents. Not because specs are bad. Specs are useful. Intent matters, acceptance criteria matter, boundaries matter. But the moment the answer becomes "write a bigger upfront spec so the agent can execute," you've recreated the exact workflow most good engineers walked away from.
And they walked away for a reason. Spec-first asks you to answer too much before you know enough. You're expected to define the full shape of the work before the code has been inspected, before the edge cases surface, before the awkward legacy behaviour is found, before the real rabbit holes show up. That's why agile, prototyping, small slices, and fast feedback won in the first place. Software is full of unknowns, and you cannot spec an unknown.
With agents the risk gets sharper, not softer. Too thin a spec and the agent fills the gaps. Too heavy and it confidently executes assumptions that were already wrong. One stale instruction, one over-confident leap, and the work spirals — faster than a human ever would.
Here's the part that gets missed: the iterative approach doesn't just protect quality at the cost of speed. I'd argue it delivers closer to what was actually wanted and gets there quicker — because you're correcting course early instead of discovering at the end that the spec was wrong. And what comes out the other side is easier to maintain, change, and scale, because it was shaped by the real system rather than an upfront guess.
So I'm not trading away engineering judgement to buy a little speed. The speed should adapt to the proven workflow, not the other way round.
You don't hand a senior developer a 40-page implementation script. You give them the outcome, the guardrails, the standards, the risks — then you expect them to inspect the system, use judgement, and bring back evidence. Agents are no different.
The future of AI delivery isn't bigger prompts or heavier process. It's just enough structure to stay safe, enough freedom to do useful work, and enough verification to know the work can be trusted.
Not vibe coding. Not AI waterfall. Disciplined, governed, iterative delivery.