Gareth Williams. Writer of Prompt to Prod and practitioner with 15+ years building software across banking, energy, government and retail. Now focused on agentic engineering, AI governance and what it takes to ship AI in production. No hype.
Notes from building AI in production - not theory, not hype.
The most important architectural decision in AI isn't technical — it's ethical. This piece argues that building human-centred AI systems, where machines handle tedious work and humans retain judgement, is both a moral obligation and a competitive advantage. The alternative — reverse centaurs where humans become accountability shields for automated decisions — is already here, and technologists must actively choose against it.
AI agents fail at complex business tasks not because the model is wrong, but because it has no machine-readable model of how your organisation actually works. This piece makes the case for enterprise ontologies — semantic layers that unify process models, data schemas and business rules into a single queryable graph — and shows why LLM accuracy on business queries jumps from 16% to 72% when one is in place.
Anthropic positions itself as the responsible steward of advanced AI, but its actions tell a different story. This piece examines the Mythos model release and what it reveals about how commercial pressure and government contracts consistently override stated safety principles — and why the concentration of critical AI infrastructure in a handful of American companies is a geopolitical problem every nation needs to take seriously.
One-off prompts are the Post-it notes of AI work — useful once, gone tomorrow. This piece applies atomic design principles to AI instructions: breaking prompts into reusable atoms (core skills), molecules (sub-workflows) and organisms (complete processes). The result is a compounding skill library where each new piece of automation makes the next one cheaper to build, and where expertise becomes tradeable rather than locked inside someone's context window.
When you grant an AI agent your credentials, you have created an employee clone that inherits all your access and none of your judgement. This piece maps the predictable trajectory of AI platforms — freemium, lock-in, monetisation, data mining — and argues that organisations must build dedicated security frameworks for AI identities now, before platform vendors decide the defaults for them. Non-human identity scoping, fine-grained RBAC and centralised AI gateways are not optional extras.
AI systems are generating code and decisions faster than the humans responsible for them can understand. Comprehension debt — the gap between what your systems do and what you actually comprehend — is accumulating silently across engineering, HR, legal and finance teams. This piece explains why it compounds like financial debt, and what deliberate practices keep it from becoming a systemic risk.
The most effective specification for an AI coding agent isn't a document — it's a working prototype with a passing test suite. This piece argues for prototype-driven development, where design teams deliver production-ready code and end-to-end tests as executable acceptance criteria. It's a fundamental inversion of the Agile model, and it's the only workflow that consistently survives contact with autonomous coding agents.
Agentic engineering — the discipline of designing, orchestrating and governing systems of autonomous AI agents — will become as fundamental to knowledge work as spreadsheet literacy. This piece maps the core practices: specification depth, orchestration patterns, quality gates and human checkpoints. The organisations that build this discipline deliberately now will compound advantages that ad-hoc AI adoption cannot match.
Agentic engineering is not a new tool in the software development toolkit — it is a replacement for the toolkit itself. This piece argues that as AI agents take over code authorship, the engineer's role shifts to architectural systems design: writing rich specifications, defining orchestration harnesses and reviewing autonomous output. The organisations that build this discipline deliberately now will hold a compounding advantage that ad-hoc AI adoption cannot close.
Bolting AI coding tools onto Scrum produces vague tickets that cause agents to hallucinate architecture decisions. This piece proposes a replacement SDLC: prototype-driven development, where designers deliver user-tested, code-based prototypes with automated end-to-end test suites before backend work begins. The prototype is not a wireframe — it is the specification, and the tests are executable acceptance criteria.
Knowledge workers lose 1.8 to 2.5 hours every day retrieving information they already have. The solution isn't more storage — it's a context engine: an organised personal system that feeds your AI tools with the right information at the right time. This piece explains how to build one without over-engineering it, and why the compounding productivity gains from a well-maintained context flywheel will define the gap between AI-native and AI-adjacent workers.
Every large language model — including GPT-4 and Claude — is built from billions of perceptrons, the binary classifiers Frank Rosenblatt described in 1958. This piece demystifies the foundational architecture of modern AI: how perceptrons learn decision boundaries, why sigmoid activation functions unlocked gradient descent, and how embeddings allow models to reason about meaning. Understanding this isn't just academic — it changes how you design systems around AI.
Code review is one of the highest-leverage practices in software engineering and one of the most consistently done badly. This piece lays out a structured approach: the right indicators to signal comment severity, how contributors can reduce friction before the first review comment lands, and why the goal is collaborative learning rather than gatekeeping. Written before AI coding agents changed the review surface — but the principles hold even more strongly now.
Technology professionals hold power comparable to physicians over the lives of millions, yet operate without any equivalent ethical framework. This 2017 piece — which reads as prescient given what followed — argues for a voluntary technologist's code of conduct: not rigid legislation that stifles innovation, but a shared commitment to accountability before deployment that the industry has consistently failed to adopt on its own.
I write and speak about what AI means for the people building software - the craft, the architecture, the governance and the risks most adoption skips. Plain, practitioner, no hype. Based in Australia; remote podcasts and talks worldwide, in-person around Melbourne. Available for first bookings.
When AI writes faster than your team can understand - the structural risk no one is pricing in.
The employee-clone problem - non-human identity and why AI agents need their own security model.
Why your AI stalls without a model of your business - and how to build one.
The craft nobody is teaching yet - what changes when you design the system that writes the code.
An SDLC for the AI age - the best spec is a tested prototype, not a document.
Copy-paste bio (50 words)
Gareth Williams is a Principal Solutions Consultant at Versent and the writer behind Prompt to Prod. He has spent 15+ years building software across banking, energy, government and retail, and now focuses on agentic engineering and AI governance. He writes and speaks about what AI actually changes for engineering teams, not the hype.
I take on the occasional advisory engagement - usually agentic engineering strategy, AI governance or getting AI systems from proof-of-concept to production. If that is useful to you, get in touch.
I've spent 15+ years building software across financial services, energy, government, eCommerce and health, for organisations including Westpac, Lloyds Banking Group and Shell Energy. I write the Prompt to Prod newsletter and speak about what AI is doing to the craft of software.
My current focus is agentic engineering and AI governance - the discipline of taking AI from a compelling demo to a system that runs reliably in production, and the organisational and ethical side of getting that right.
Based in Australia. Working globally.
I'm open to podcast appearances, speaking, guest writing and the occasional advisory engagement. Tell me what you're working on.
I typically respond within one business day.