The AI-Native Delivery Loop
Coding agents can shrink the build step from days to hours while the delivery calendar barely moves, because typing was never the slow part. This eight-part series sets out an AI-native delivery loop: intent written first, the specification governed as the control surface, one living specification instead of restated handoffs, the harness treated as a control system, acceptance authority kept with named people, AI code review used as a filter rather than a sign-off, production signals routed back into the artefacts agents read, and AI used in the product only where it beats plain code. One fictional Cape Town payments company and its merchant onboarding squad carry the examples, and every figure is illustrative.
8 articles
- Part 1AI-native delivery is not faster deliveryCoding agents can shrink the build step from days to hours, and the delivery calendar barely moves. Part 1 explains why: the slow parts of delivery were never typing, and an AI-native lifecycle redesigns the loop around intent, specification, evidence and production feedback rather than speeding up one step.8 min read
- Part 2The specification is the control surfaceOnce agents read the specification directly, it stops being documentation and becomes the main way you steer their work. Part 2 covers what that demands: versions, owners, explicit boundaries and a way to tell every task, test and agent when the rule they were built on has changed.7 min read
- Part 3One living specification beats twelve stale handoffsEvery role in delivery keeps its own version of the truth: the business case, the analyst's notes, the architecture document, the tickets, the test plan, the runbook. Each restates the one before and loses a little. Part 3 shows why agents make that drift more expensive, and how one owned, versioned specification replaces the chain.6 min read
- Part 4Harness engineering has grown up: from four artefacts to a control systemWe have already explained the coding-agent harness as guides, guards, checks and sensors. Part 4 shows what the newer control-system framing adds: who owns which layer, the cheapest reliable control first, three things a harness must regulate, why some codebases are easier to harness than others, and the steering loop that keeps it improving.10 min read
- Part 5Agentic development still needs an acceptance authorityAn agent can write the code, run the tests, fix what failed and report that it is done. None of that decides whether the work is the right work, whether a risk is acceptable or whether the change may reach customers. Part 5 separates execution, which can be delegated, from acceptance authority, which cannot.6 min read
- Part 6AI code review is a filter, not a sign-offAI code review tools find real problems quickly, and they are now cheap enough to run on every change. Used well, they are a filter that improves what reaches a human. Used badly, a clean AI review becomes an approval nobody actually gave. Part 6 sets out where automated review belongs and what stays with people.6 min read
- Part 7Production is part of the promptAgents build from what they are given: intent, specification, architecture rules and the harness. If production behaviour never changes those inputs, every new cycle starts from assumptions production has already disproved. Part 7 shows how to route logs, incidents, user behaviour, cost and quality signals back into the artefacts agents read.6 min read
- Part 8AI-native does not mean AI everywhereAn AI-native delivery loop uses agents to build software. It does not follow that the software should use AI at every step. Part 8 closes the series with a simple method for deciding where deterministic code, ordinary automation, AI and human judgement each belong, and why most of a well-built system is still plain code.6 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala