Insights
Field notes on the future of software delivery.
How we think about the decisions AI forces. Guided perspectives by role, and the full library by theme — pick a view; it remembers where you left off.
Persona doors
“I’m an X — where do I start?”
Two ways in
Perspectives
Curated routes through the work, by role. Start where your question is.
Making better AI decisions
You do not need to become an AI expert to govern AI well — only the skills you already use to allocate capital, manage risk and decide under uncertainty. This is how to apply them.
3 articles · 36 minBegin →Recovering the discipline of choice
More ideas can now be built than any organisation could possibly absorb. This is about knowing which ones deserve attention, funding and leadership energy — and which to stop.
5 articles · 55 minBegin →Who creates the next senior
Every organisation worries about replacing early-career work. Very few are asking who replaces the journey that created their seniors. This is about rebuilding that on purpose.
2 articles · 20 minBegin →Why everything feels different
If AI feels exciting, confusing and impossible to ignore all at once, start here — what actually changed, what didn’t, and why a career’s worth of judgement matters more now, not less.
6 articles · 46 minStart the foundation →Strategy, economics and governance of agentic delivery
A board-level model: what changes, how to measure it, what it costs, and how to keep it safe.
10 articles · 92 minBegin →Architecture as decision-making in the agentic era
Treat architecture as the residue of decisions — and design systems that stay changeable.
8 articles · 72 minBegin →Product delivery that survives contact with AI
Why transformations fail, and how to run delivery around acceptance, not output.
6 articles · 45 minBegin →Before you spend more on AI
A clear-eyed read on cost, value and the judgement you quietly spend down as you automate — so you adopt on evidence, not enthusiasm.
7 articles · 63 minBegin →Running delivery when agents do the building
How cadence, ceremonies, estimation and team roles change once acceptance — not capacity — is the constraint.
6 articles · 44 minBegin →Designing organisations for the agentic era
The bridge between deciding and building: how team shape, cadence and delivery operations are redrawn when the unit of work becomes a human plus an agent fleet.
5 articles · 44 minBegin →Agentic Engineering
Moving teams beyond vibe coding to disciplined, agent-assisted delivery. Operationalised by the The AI Engineering Maturity Model framework.
What Is Agentic Engineering?
Once generation becomes abundant, the constraint moves to acceptance. That single shift reorganises how software gets built.
AI Engineering · 6 minBeyond Vibe Coding
Vibe coding optimises the step that just became abundant and starves the one that just became scarce. Fine for exploration. Dangerous as an operating model.
AI Engineering · 6 minThe Agentic SDLC Is an Acceptance-Gate Problem
AI now reaches across the whole lifecycle, not just the IDE. The design question is no longer where agents act, but where a human must accept, and what evidence the agent owes them at each gate.
AI Engineering · 7 minAgentic Code Review
Reviewing AI code is not reviewing human code. The author cannot tell you where it was unsure, so the gate that closes the Acceptance Gap needs a different checklist, not the old one applied faster.
AI Engineering · 6 minContext Engineering: Context Is the New Architecture
Most context-engineering advice optimises the ephemeral window. The durable asset is the context layer you own and version: the conventions, decisions and contracts that let a generated change be judged correct, fast.
AI Engineering · 7 minThe Memory Hierarchy of an Agentic Team: Choosing Between Convention Files, Retrieval, Graphs and Fine-Tuning
Context engineering has quietly become a systems-design problem with four competing substrates. Stop picking one by default; design a memory hierarchy with eviction, freshness and provenance rules.
AI Engineering · 6 minChoosing Models for Engineering Teams
The leaderboard is the wrong place to start. Once code generation is abundant, the model is a commodity input and the harness is the product, so model choice rarely closes the acceptance gap. Choose by use case, cost-at-acceptance, verified context and the data boundary instead.
AI Engineering · 9 minThe AI-Native Engineering Team
Once acceptance is the constraint, the org chart is wrong. Four judgement roles must be owned on every team — and 'more humans' now hurts.
Platform · 9 minEnforcing Determinism in Probabilistic Systems
You cannot make a language model deterministic — and chasing that is the wrong goal. Make the system deterministic at the boundaries that matter: generate freely, validate strictly, and ship only what a gate would accept.
AI Engineering · 7 minThe Eval Is the Spec: Why Acceptance Criteria Become Executable Tests in Agentic Delivery
When agents generate code for free, the prompt stops being the binding artefact. The evaluation harness becomes the real specification - and the team that writes the best evals, not the best prompts, controls quality and velocity.
AI Engineering · 8 minThe Integration Seam Is Where AI-Generated Software Breaks: Payments, Identity and the Limits of Generation
AI agents write clean code inside a service and confident nonsense at the boundary between systems. Idempotency, payment state machines, token refresh races and eventual consistency live in vendor quirks and production incidents, not in training data — and that is exactly where acceptance must now concentrate.
Architecture · 6 minEvent Contracts as the Coordination Layer for Mixed Human and Agent Teams
When some of the work is done by people and some by autonomous agents, the durable coordination mechanism is not the org chart or the ticket. It is the versioned, well-owned event contract — and schema governance quietly becomes agent governance.
AI Engineering · 9 minPrivate by Architecture: Running Local and Self-Hosted Models When Code, Payments or Identity Data Cannot Leave
For teams handling cardholder data, PII or proprietary source under regulatory constraint, the model-selection question is not which frontier model is smartest but where inference can legally and architecturally run. A deliberate tier of local and self-hosted models — with the capability gap closed by tooling, retrieval and orchestration rather than raw model quality — is now a practical engineering choice, not a compromise. Design the data boundary first, then place models inside it.
AI Engineering · 6 minMeasuring AI Engineering Properly
Lines of code and "percent faster" were always vanity metrics. Now they are actively misleading. If generation is abundant, measure the scarce step: acceptance, rework and stability.
Trust, Governance & the Economics of AI
How machine-produced work earns the right to ship — and what it costs. Operationalised by the The Governance-to-Value Ratio framework.
Designing the Human Acceptance Loop: Where People Must Stay in the System as Agents Take Over the Build
As agents absorb the build, the decisive human work shrinks to a few acceptance moments. Keep humans in every loop and you lose the leverage; remove them and you lose accountability. The discipline is to design the loop on purpose.
AI Engineering · 7 minAI Coding Governance That Enables, Not Forbids
Most AI coding policies are written to stop something, and they fail. The governance worth building does the opposite: it widens safe adoption while raising acceptance on evidence, not enthusiasm.
Delivery Assurance · 6 minProvenance Engineering: Reconstructing Who Decided What When Humans, Agents and Models All Contributed
When software is co-produced by humans, agents and models, "who decided this, on what basis, and can we reconstruct it?" stops being a forensic luxury and becomes a first-class engineering requirement.
AI Economics · 7 minThe AI Delivery P&L: Why Generation Got Cheap and Your Cost Base Did Not
Boards modelling AI savings on developer headcount are budgeting the cheapest part of the problem. When generation costs nothing, the bill moves downstream — to acceptance, integration, review and ownership — and the evidence says it gets bigger.
AI Economics · 8 minBuild, Buy or Generate: How Agentic Delivery Rewrote the Oldest Capital-Allocation Decision
Agentic generation collapses the cost of building — but it does not tilt build-versus-buy toward build. It dissolves the decision's premises and adds a third path with its own ownership liability.
Product Delivery · 7 minDelivery Telemetry: Instrumenting the Path from Intent to Production So You Can See Where It Stalls
Organisations instrument their running systems to the millisecond yet leave the journey from business intent to deployed outcome almost entirely un-observed. Once AI accelerates the build, the constraint moves upstream into decision and acceptance, and only instrumentation will show you where it went.
Delivery Architecture
Preserving intent from strategy through product to running outcomes. Operationalised by the Delivery Architecture: The Translation Layer framework.
The Missing Architecture Layer Between Strategy and Delivery
Strategy and delivery both have owners. The translation between them — capability maps, domain models, target and transition states — has none. Name it, and call it Delivery Architecture.
Product Delivery · 6 minWhy Most Product Transformations Fail Before Engineering Starts
The failure is decided before the first sprint is planned. By the time engineering inherits the work, the objectives are already ambitions, discovery has been treated as a phase to clear, and stakeholders are aligned on slogans rather than trade-offs. A team that builds fast simply reaches the wrong destination sooner.
Product Delivery · 7 minProduct Delivery in the Age of AI
AI now touches discovery, requirements, planning, build, test and docs. It compresses execution everywhere — and improves intent-translation nowhere. When generation is cheap, acceptance becomes the constraint.
Product Delivery · 7 minFrom Idea to Production: A Practical Product Delivery Lifecycle
Most delivery teams pour their attention into Build — the one stage the evidence says is least correlated with success. A practical, nine-stage lifecycle for shipping value, not volume.
Product Delivery · 7 minMVPs, Pilots and Production Systems: Knowing the Difference
Prototype, MVP, pilot and production are not sizes of the same thing — they are four different questions, each with its own acceptance bar. Confusing them causes over- and under-engineering, which are the same mistake.
Product Delivery · 7 minMeasuring Product Delivery: Beyond Velocity and Story Points
Velocity and story points measure how busy engineers are, not whether business intent became a sustained outcome. A field guide to building a delivery scorecard as a translation audit, where speed is never celebrated without acceptance.
Product Delivery · 8 minProduct Delivery Governance Without Bureaucracy
Heavyweight approval boards slow delivery without making it safer. The fix is not less governance but governance designed around decision quality: separate the irreversible decisions that deserve scrutiny from the reversible ones that just need to ship.
Product Delivery · 8 minBuilding Product Teams That Scale
Scaling a product team is an intent-translation problem, not a headcount problem. Why the org chart is a lagging artefact, and a decision-ownership migration map for the 10-30-50-150 thresholds.
Product Delivery · 7 minWhy Alignment Beats Agile: Product, Architecture and Engineering
Most teams are Agile. Few are aligned. Product describes one system, architecture models a second, engineering ships a third. The cadence improved; the seams stayed un-owned. Alignment is the real constraint.
Architecture & Decisions
Architecture as decision quality and velocity, not diagrams. Operationalised by the Decision Architecture framework.
Architecture in the Age of AI
When agents become co-authors, architecture stops being about diagrams and becomes the discipline of decisions that let an organisation accept changes it did not write by hand.
Architecture · 7 minArchitecture Is Not About Technology — It Is About Decision-Making
Diagrams and stack choices are the visible residue of architecture, not the work itself. The work is making trade-offs under uncertainty — choosing constraints and accepting risk for a specific context with incomplete information. The expensive failures are not bad pictures; they are good decisions nobody recorded and poor ones nobody could reverse.
Architecture · 7 minArchitecture Decision Records: The Missing Link Between Architecture and Delivery
Teams document what they built, not why. The rationale — the rejected options and the trade-off accepted — is the actual deliverable, and increasingly the context layer your AI agents are missing.
Architecture · 8 minWhy Most Architecture Functions Fail to Create Business Value
Architecture teams drift into review boards, document factories and gatekeepers. Their actual job is the opposite: reduce uncertainty, accelerate delivery and enable decisions. The evidence on why the drift destroys value is now hard to ignore.
Architecture · 7 minEvent-Driven Architecture Beyond the Technology
The broker was never the hard part. Once you have chosen Kafka or a queue, the difficult ninety per cent remains: who owns the event, what it means, and who is paged when it silently lands in a dead-letter queue.
Architecture · 7 minArchitecture Governance Without Bureaucracy
Heavyweight approval boards are empirically weak and quietly expensive. The alternative is not chaos — it is decision records, an advice process, and governance tiered by consequence rather than hierarchy.
Architecture · 7 minDesigning for Change: The Most Underrated Architecture Principle
Great architecture is not optimised for today's requirements, the one set you already understand. It is optimised for the changes you cannot yet name. But designing for change does not mean maximising flexibility everywhere; that is its own failure mode. It means deciding which changes you will make cheap to absorb, which decisions are genuine one-way doors, and engineering reversibility into the rest.
Architecture · 8 minArchitecture as a Product: Designing Platforms People Actually Use
Internal platforms fail when they are designed for the people who govern them rather than the people who use them. The smoking gun is a perception gap: producers are convinced a mandatory platform works while consumers are split. Adoption, freely given, is the only honest metric, and a mandate is an admission the architecture could not win it.
Product Delivery · 8 minThe Architecture Decisions That Determine Product Success
Build vs buy, platform, integration, scalability, domain boundaries — five 'technical' choices that are really bets on what the business will become. Architecture is where intent survives translation, or quietly dies.
Architecture · 7 minThe Architecture Decisions That Matter Most in Enterprise Transformation
Build/buy, platform versus product, integration, identity, and data ownership are not procurement line items. They are the one-way doors that decide whether a transformation succeeds before delivery even begins.
More insights
Platform Engineering for Agentic Teams
When AI agents become contributors, the internal platform stops being plumbing and becomes the control plane that decides whether AI compounds your strengths or your dysfunction.
Platform · 9 minGolden Paths: Preserving Intent at Platform Scale
A golden path is an architecture decision that only has to be made once — and, in the agent era, the only scalable way to give every contributor, human or not, the same intent.
Platform · 9 minCognitive Load, Not Feature Count
The best thing a platform — or an AI — can do is let a good engineer not think about something. That, not features shipped or lines generated, is the number worth chasing.
AI Engineering · 12 minFrom Tasks to Roles
Agents execute tasks. People hold roles. They are not the same unit — and confusing them is the quiet mistake under most AI-engineering failure. A three-layer model for what actually changes, and the part of the work that has a floor.
AI Engineering · 9 minThe Accountable Core
Strip away every task an agent can run and every role it can assist, and something remains that cannot be handed over: the answerability for the outcome. Delegation has a floor — and that floor is where seniority now lives.
AI Economics · 9 minThe Missing Rung
Seniors learned the judgement they now hold by doing the early-career work we are handing to agents. Automate the first rung of the ladder and you may quietly remove the rung that made the top — a stock of judgement we are spending down before we can measure it.
Product Ventures · 11 minDeciding What to Build When Building Is Cheap
When AI makes building cheap, the cost of building the wrong thing does not fall with it — it compounds. Most ideas never move the needle, so abundance just ships the duds faster and louder. The scarce, decisive skill stops being construction and becomes judgement: the discipline to decide what not to build. We call the gap that opens the Conviction Gap.
Transformation Leadership · 10 minWhy AI Pilots Stall Before ROI
Most AI pilots work — and still return nothing — because they are built to prove the technology, then handed across a last mile no one owns into an operating model that never changed. Adoption is not value. We call the chasm between a working pilot and a changed operating model the Production Gap.
Product Ventures · 11 minThe Unit Economics of an AI Product
Classic software was almost free to run; an AI product pays real money every time a customer uses it. Inference behaves more like cost of goods sold than infrastructure — and falling token prices do not automatically fix margins, because consumption expands to absorb the savings. AI changed the economics of software, not just the mechanics. We call the line you now have to design for the Margin Floor.
Transformation Leadership · 11 minWhy Most AI Use Cases Should Never Reach Production
Most AI failures are not technical failures — they are governance failures disguised as technical programmes. Most ideas simply should not ship, and the real risk is not deploying the wrong thing but being unable to stop it. The highest-leverage AI capability is the ability to stop: Default Off as the philosophy, the Kill Rate as the measure.
Transformation Leadership · 12 minThe Decision Architecture of AI
Organisations spent decades architecting systems and almost no time architecting decisions. The missing layer in enterprise AI is not model quality — it is decision architecture: who decides, on what evidence, and how reversibly. Doing nothing is still a decision; it is simply one you let emerge by default.
Transformation Leadership · 11 minRebuilding the Apprenticeship
The work used to teach judgement for free — bundled invisibly into the foundational work AI now automates. So judgement has to be taught on purpose: a designed apprenticeship, aimed at judging rather than only producing, with the senior’s hidden reasoning made visible. Judgement was never free; the work was paying for it.
Delivery Assurance · 11 minThe Hard Part Is Access, Not Intelligence
Intelligence without access is a demo; access without control is a breach. Most organisations are treating AI as a model problem when it is rapidly becoming an identity one — and your biggest identity problem is no longer human. The hard, durable work is the Agent Gateway: scoped identity, action-level authorization, audit and reversibility.
Transformation Leadership · 5 minThe Shift
For thirty years the hard part of software was making it. That has quietly stopped being true. Technology is becoming easier to generate; judgement is becoming more valuable — and the experience you already have is the asset, not the liability.
Context Architecture · 13 minRetrieval Architecture That Doesn’t Rot
Most retrieval systems fail long before they break. Retrieval doesn’t fail loudly — it rots quietly: the index goes stale, permissions drift, quality regresses with no error in the logs. The durable architecture is not the vector store; it is the lifecycle. Retrieval is a living system, not a launch.
Context Architecture · 13 minThe Context Supply Chain
We hardened the build-time supply chain and left the runtime one wide open. The information actually feeding an agent often arrives with unknown origin, freshness and integrity — and your agent is only as good as its worst context supplier. Context deserves a supply chain: provenance, freshness, integrity.
Transformation Leadership · 13 minThe Board-Level Narrative for AI Delivery
Hype and fear look like opposites. In practice they are often the same thing: decisions made without sufficient evidence. A board’s job is not to approve an AI strategy — it is to govern a portfolio of reversible, evidence-graded decisions. The boardroom is the highest tier of decision architecture.
AI Engineering · 8 minRole Fluency
When agents take the tasks, the engineer’s core skill stops being execution and becomes the ability to move cleanly between roles — decomposer, reviewer, arbiter, owner — and to know, at any moment, which one they are in. Most bad agentic work is a role error.
Delivery Assurance · 10 minSecurity Review for AI-Generated Code
AI writes code that compiles about 95% of the time and is secure barely half the time — and that gap has not moved in two years. The review that matters is no longer “does it run?” but “what did it quietly let in?”
Delivery Assurance · 9 minWhen an Agent Ships a Regression
Generation is cheap, so more change reaches production faster — and some of it is wrong. The hard question is no longer who to blame, but how you detect, contain and recover when the change’s author was a machine.
Delivery Assurance · 9 minThe Audit Trail for Agentic Delivery
When humans, agents and models all touch a change, “who decided this?” stops being obvious. Provenance can prove how an artifact was built; it does not yet prove who — or what — decided it. That gap is where accountability has to be engineered.
Product Delivery · 9 minScrum in the Agentic Era: Which Ceremonies Survive
When agents write much of the code, the question is not whether Scrum dies. It is that its ceremonies were tuned for scarce engineering capacity — and capacity is no longer the constraint. Re-point them at the one that is: acceptance.
Delivery Assurance · 9 minContinuous Assurance: From Audit Day to Always-On Evidence
When agents ship change continuously, a point-in-time audit is obsolete the morning after it passes. Assurance has to become a signal generated at machine speed — control-as-code in the pipeline, not a certificate on the wall.
Product Delivery · 9 minThe Agentic Org Chart: Team Shape When Your Teammates Are Agents
Beyond who does what, agents change how a team is shaped — its size, its boundaries, and Conway’s Law when part of the communication structure is autonomous. The surprise: Team Topologies survive the transition, and matter more.
Platform · 10 minCI/CD, DevOps, DevSecOps, FinOps: The Convergence into Agentic Delivery Operations
Four disciplines grew up separately. When agents author most changes, they collapse onto the same control point — the merge gate — and become one thing: agentic delivery operations.
AI Engineering · 9 minThe AI Productivity Paradox
Every developer feels faster. The release cadence has not moved. The gap between individual speed and delivered outcomes is the defining measurement problem of agentic engineering — and it has a cause you can fix.
Delivery Assurance · 8 minSlopsquatting: When Your Agent Invents a Dependency
AI coding assistants confidently import packages that do not exist — and attackers have learned to register the names before you do. It is a new, predictable supply-chain attack, and the fix belongs at the merge gate, not the quarterly scan.
Delivery Assurance · 9 minThe EU AI Act Meets Your Delivery Pipeline
On 2 August 2026 a new tranche of the EU AI Act takes effect — even as Brussels debates delaying part of it. For teams building software with and into AI, here is what actually changes, and what to do now without waiting for the dust to settle.
AI Engineering · 9 minCognitive Debt: The Maintainability Bill for AI Code
AI lets you ship code faster than anyone can understand it. The gap between what runs in production and what your team comprehends is a new kind of debt — and it compounds quietly until the day no one can safely change the system.
AI Engineering · 8 minWhy Senior Engineers Slow Down With AI
The counterintuitive finding of the last two years: AI helps newcomers most and experts least — sometimes making the most experienced engineers slower. Understanding why explains where AI actually pays off, and where it quietly taxes your best people.
Transformation Leadership · 8 minAgent Washing: Telling Real Agents From Rebadged Automation
Everything is suddenly “agentic”. Most of it is a chatbot bolted to a workflow you already had. Here is how to tell a genuine agent from a relabelled one — before you buy, build, or stake a roadmap on the difference.
Transformation Leadership · 9 minHow to Evaluate an Agentic-AI Vendor: A Due-Diligence Checklist
The demo always works. The questions that predict whether an agentic-AI product survives contact with your business are the ones vendors are least prepared for. Here is the checklist to bring to the table.
AI Engineering · 9 minOne Agent or Many? When Multi-Agent Helps and When It Hurts
Multi-agent architectures are the fashion of 2026 — and frequently the wrong call. Coordination buys you parallelism and specialisation at the price of new, harder failure modes. Here is how to decide which you actually need.
AI Engineering · 9 minObservability for Agentic Systems
Traditional monitoring tells you the service is up. It cannot tell you why an agent decided to do what it did. When the actor is non-deterministic and acts through tools, you need to trace intent to action to outcome — or you are flying blind.