The enterprises unlocking real productivity from AI coding agents aren’t the ones with the best tools. They’re the ones that recognise the cultural shift required to bring agents into their teams.
I recently published a deep-dive on InfoQ covering how Spec-Driven Development can be adopted at enterprise scale — the tooling gaps, brownfield integration patterns, multi-repo orchestration, and the long-term shift toward spec-native delivery. If you haven’t read it yet, I’d encourage you to read the full article on InfoQ.
This post includes ideas that complement that article — ones that feel increasingly urgent as more organisations start putting AI coding agents into their engineering workflows.
Onboarding a Coding Agent Is Not Free
The framing around AI coding agents is a little too over-simplified: you simply start using them and suddenly you can expect to see major productivity boosts. It rarely works that way.
Think about the last time you brought a new engineer onto your team. The first few weeks are expensive. You invest time explaining the codebase, the team conventions, the unwritten rules, the business domain, the history of why things are the way they are. The context you share with them is not wasted — it is the thing that makes everything after it faster.
Adding an AI coding agent carries the same tax, except most teams don’t pay it upfront. They spin up the agent, give it a task, and then spend the next few hours cleaning up the debris. That’s not the agent’s failure. That’s a context engineering failure.
Context engineering is the discipline of giving an AI coding agent what it needs to know — your architecture, your conventions, your boundaries, your constraints. It’s the functional equivalent of onboarding a new joiner. And just like onboarding, it requires intentional effort. You have to decide what the agent needs to understand. You have to maintain that context as things change.
This is precisely where SDD earns its place. A well-structured spec gives a coding agent the same thing a good onboarding gives a new hire: the business context behind the work, the technical approach that fits your architecture, and a clear breakdown of what needs to happen and in what order. Hoping that better models will solve the problem is not a strategy — because the problem was never the agent. It is the absence of context.
The Culture Shift Nobody Talks About: From Instructions to Conversation
There is a deeply ingrained habit in how most engineers interact with AI today: they issue instructions. Write this function. Fix this bug. Refactor this class. The model complies, the human reviews, and the cycle repeats.
This is a perfectly valid way to use a calculator. It is a poor way to use a collaborator.
The shift that distinguishes teams getting outsized value from AI coding agents is the move from instructional to conversational interaction. Instead of asking the agent to produce an output, they use it to think through the problem first — to surface edge cases, challenge assumptions, and explore alternatives before any code gets written. The spec that comes out of that conversation is richer and more considered than anything a single prompt would produce.
This requires engineers to change how they think about their role. If your career has been defined by the ability to produce — to ship code, close tickets, hit velocity targets — slowing down to invest in articulation before execution can feel counterproductive. It takes intentional practice, and it requires organisations to signal that this kind of thinking is valued. An agent given well-reasoned, clearly articulated intent will do better work across more tasks, for longer, with less correction than an agent that received a crisp one-line instruction.
What makes SDD the structural foundation for this shift is that the spec becomes the shared interface — the place where product, architecture, and engineering converge to build execution context together, rather than passing handoffs down a chain. Organisations that build this culture — making space for deliberation, investing in spec quality, treating intent articulation as a senior engineering skill — will be the ones that actually unlock what AI coding agents are capable of.
The full technical picture — how SDD tools support this, where the gaps are, what adoption looks like in practice — is in the InfoQ article. What I wanted to say here is simpler: with AI agents capable of sustained autonomous execution, the bottleneck has shifted from how fast we write code to how effectively we articulate intent. SDD is how we rise to meet that shift — but only if we treat it as an organisational capability to develop, not just a technical practice to install.