- Telemetry across more than ten thousand developers shows 98 percent more pull requests and 91 percent longer review times with flat delivery metrics. Specification discipline, not model capability, is the binding constraint.
- About two in ten organizations are scaling coding agents, and a third have built software they would otherwise have bought, so the volume of software to maintain is going up.
- Three drivers compound: delegation becomes normal, work becomes committed artifacts, governance moves out of meetings into enforcement.
- Six consequences for 2027: build is not the constraint, work moves as artifacts, governance lives in hooks and branch protection, a hand-written system model is essential, review capacity is the ceiling, tests and evals replace line-by-line review.
- The open question is who operates the loop. Tuesday 8 December, 1 PM ET: the 2027 condition, the loop, one real change run live, and ten questions for any vendor or plan.
Build went from weeks to hours. The process did not move. This is the argument behind our 8 December session on what the AI-native SDLC becomes in 2027, and who runs it.
Build went from weeks to hours. The process did not move.
That is the condition most engineering organizations are in as 2026 closes, and it is the reason we are running a session on 8 December called What does AI-native SDLC look like in 2027? This post is the argument behind it: what is measurably shifting, what is driving it, and what it means for anyone who has to plan an engineering budget for next year.
The shift, concretely
In 2024 an AI coding tool was an autocomplete with ambition. In 2026 an agent takes an issue, writes the code, runs the tests and opens a pull request, and the engineer’s job is to decide whether that pull request should exist.
The volume has changed before anything else has. A May 2026 paper on arXiv that reviewed 67 studies from 2022 to 2026 put it bluntly: telemetry across more than ten thousand developers shows “98% more pull requests but 91% longer review times with flat delivery metrics.” Controlled studies find productivity gains of 20 to 56 percent on well-scoped tasks; the most rigorous randomized trial found a 19 percent slowdown for experienced developers on their own codebases. The paper’s name for this is the Productivity-Reliability Paradox, and its conclusion is the sentence to carry into 2027: “Specification discipline, not model capability, is the binding constraint on AI-assisted software dependability.”
Let’s be precise about what that says. The models are not the bottleneck. The process around them is.
The evidence that it is moving anyway
The organizations scaling this are not waiting for the paradox to resolve. McKinsey’s 2026 State of AI survey, published in August across 1,719 respondents, found about two in ten organizations scaling software coding agents, rising to 31 percent at enterprises above a billion dollars in revenue. Nearly a third, 32 percent, reported deciding against buying at least one software product because they could build it internally with agentic coding tools.
Read that second number again. The build-versus-buy line is moving, which means the volume of software each organization owns and must maintain is about to go up, not down. Every one of those built-not-bought systems needs a lifecycle.
On the vendor side, Anthropic published its AI-Native SDLC Playbook in August: twelve plays across the lifecycle and a governance layer underneath them. It is a vendor’s description of the practice, and we treat it as a primary source for that, nothing more. What matters is that the practice now has a written shape. People are no longer arguing about whether the lifecycle changes, only about how far and who runs it.
What is driving it
Three forces, and they compound.
Delegation is becoming normal. Teams moved from “AI suggests, human writes” to “human specifies, AI executes, human reviews.” Once a task can be delegated end to end, the question of how many tasks to delegate becomes an economic one, and economics only push one way.
Work is becoming artifacts. A ticket describes an intention. A committed chain of artifacts, intent to specification to plan to pull request, is something an agent can execute against and a reviewer can audit. Organizations that make the switch find their velocity; organizations that keep tickets find their agents guessing.
Governance is moving out of meetings. The review that used to happen in a Thursday change board now has to happen at the speed of the loop, which means it happens in hooks, branch protection, tests and evaluation suites. Governance as enforcement rather than governance as attendance.
What it means for you in 2027
If the loop is where software gets built, six things follow, and the session on 8 December walks through each of them with a live change running end to end.
- Build is no longer the constraint; human-speed stages are. Requirements, review and decision are the new critical path. Plan capacity there.
- Work moves as committed artifacts, not tickets. If your backlog is prose, it cannot be delegated.
- Governance moves from meetings into enforcement. Hooks, branch protection, gate records. If a rule is not enforced in the pipeline, it is advice.
- A system model written by hand is essential. The agents need to know what the system is, and that file does not write itself. We have written before about what your codebase knows that your team forgot.
- Review capacity, not seat count, is the ceiling. You can buy more agents. You cannot buy more senior judgment by the seat.
- Tests and evaluations become controls that replace line-by-line review. Evidence, not eyeballs, decides what ships.
None of these is a tool decision. All of them are operating-model decisions, which is why the AI SDLC literacy conversation moved from tools to stages this year, and why the practical question underneath is the one in the title: run the loop yourself, or have it run for you.
How to position for it
Here’s a pragmatic path for the next quarter:
- Pick one system and write its model down. Architecture, boundaries, the rules nobody wrote. If it takes a week, that week is the cheapest one in your 2027 plan.
- Convert one class of work, bug fixes say, into an artifact chain and run it through an agent with a named gate at the end. Measure review time, not build time.
- Decide, in writing, what an agent may do and may not do without a human. Put that in branch protection, not in a wiki.
- Then decide who operates it. In-house with staffing you do not yet have, or operated for you with your engineers supervising the gates. That is the decision How to Build an AI Factory ends on, and it is the one we spend the last ten minutes of the webinar on.
Tuesday 8 December, 1:00 PM ET / 6:00 PM UK, 60 minutes with live Q&A. Nick Chase and I will cover the 2027 condition, what the loop contains, one real change run through it live, and ten questions you can put to any vendor, outsourcer or internal plan, including ours. Register free.
Sources: Farrag, “The Productivity-Reliability Paradox: Specification-Driven Governance for AI-Augmented Software Development,” arXiv 2605.01160, May 2026. McKinsey, “The state of AI in 2026: On the road to ROI,” August 2026. Anthropic, “The AI-Native SDLC Playbook,” August 2026.
