Writing archive
Older essays on agentic systems, governance, operations, incentives, and the work of making AI useful in real organisations.
Search articles · Choose a reading route
We built the physical world around human hands and the digital world around human clicks. Browser agents are an ingenious bridge, but the future belongs to structured tools and interfaces designed for human oversight.
Part three of Agentic Operating System For Your Business: choosing between the deep integration of an agent harness and the flexibility of portable API workers.
Use the agent harness for leverage today, but keep a tested API or local-model route so convenience does not become operational dependency.
The July 2026 UK baseline for direct AI API prices, showing quick, general, and frontier models alongside reference workload costs and the qualifications that affect a completed task.
A personal note to Alex about moving from ChatGPT on a phone to a simple Restaurant Control Room in Codex on a laptop.
A thought from OpenAI Build Week in London about uncertainty, pricing, and building where no reliable pattern exists yet.
A practical experiment with Focus Reader, dyslexia, attention, and comprehension.
Before you give an agent more intelligence, decide what failure you can survive. Operational resilience is the first task, not the final check.
AI can advise, challenge, and inform. It cannot take responsibility for the judgement you choose to make.
AI can help us build at extraordinary speed. It cannot reliably warn us about rules, risks, and details we do not yet know exist.
Calling an AI tool an employee does not make it ready for work. A practical safety guide to onboarding agents through bounded tasks, context, supervision, tests, least privilege, and earned authority.
Why Tony is replacing message streams and human issue lists with bounded work packets that agents can triage, discuss, approve, execute, and receipt.
A plain-English guide to OpenAI's GPT-5.6 model family, ChatGPT, ChatGPT Work, Codex, the API, reasoning settings, tools, skills, MCP connectors, and multi-agent work.
Part two of Agentic Operating System For Your Business: creating a short AI-voiced podcast was straightforward, but publishing it exposed the identity, delegated authority, permissions, and audit layer that agent-ready platforms need.
A practical thought piece on why vibe coding changes the buy-vs-build equation but does not remove the cost of human time, maintenance, security, backups, data integrity, audit, and ownership.
A researched thought piece on why AI may lower the cost of writing code, but does not remove the need for engineers who understand architecture, security, data, scaling, reliability, operations, and business constraints.
A practical thought piece on why better frontier models will make standard operational work faster and more similar, pushing differentiation toward human creativity, taste, context, customer-specific imagination, and new creative tooling.
A documentary-style opening article for the Agentic Operating System For Your Business series, following the Chief Agentic Officer Briefing from zero useful corpus into a live agentic business experiment.
A researched thought piece on a suggestion from Athena, who works with Tony Wood in Agentic Research, that Esperanto can inspire a neutral, precise, human-readable auxiliary register for agent-to-agent communication.
A researched thought piece on why agents can improve logistics, memory, matching, research, and follow-through, while physical communities remain vital for trust, judgement, tacit knowledge, belonging, challenge, and formation.
A practical thought piece on why cron jobs, smart task lists, and fixed AI workflows are useful but not the same as genuinely agentic systems with loops, state, recovery, and goal-directed continuation.
A practical opinion piece on Garfield AI, small claims, access to justice, and the right balance between AI-supported legal preparation and human legal accountability.
A technique-led essay on goal-based agentic loops: why the human role shifts from step-by-step controller to shepherd when agents have enough context, constraints, feedback, and stop-lines to work toward a clear outcome.
A personal opinion piece on frontier AI access, operational resilience, sovereign AI, and why open models matter when the best hosted tools can be restricted suddenly.
Privacy choice
Cookieless Matomo analytics help me understand which public pages are useful. They are on by default unless you turn them off, and your choice never affects access to the site.