Blog

Notes from the workshop

What I'm learning, building, and breaking. The thinking behind the systems I ship.

All posts ai-agents9 ai-native-dev9 otagelabs8 agent-orchestration5 claude-code4 bootstrapping4 software-craft4 developer-experience3 multi-agent3

2 June 2026· 3 min

The recorder that wrote its own blog post

John Allsopp pulled a recorder out of his pocket at the AI engineer conference in Melbourne. One week of interleaved building later, it delivers meeting summaries as open markdown. The device recorded the ideation session for its own blog post.

hardwareai-agentsfirmware

18 May 2026· 6 min

The shakedown

$1,500 in tokens on a Tuesday, and a $10 subscription covering the tab. Anthropic is repricing inference, Uber blew its annual AI budget by March, and GitHub killed flat fee Copilot pricing because agentic workloads broke the model. The cost of inference is the elephant in the room, and the shakedown is coming for every business model pretending it doesn't exist.

inference-economicsai-agentsanthropic

12 May 2026· 5 min

Finding the knack amongst all the tools

A broken car in Metung, a garage full of tools, and no idea where to start. That is what an AI agent session looks like from the inside. The tools exist; the knack for using them does not come installed.

ai-agentsmcpcontext-engineering

6 May 2026· 4 min

Beware there be dragons when sending your AI into the jungle

Two production databases deleted by AI agents in two weeks. The incidents tell different stories, but they point to the same gap: if nobody in the room knows where the dragons are, the AI will walk straight into the cave.

ai-agentsai-safetydevops

30 April 2026· 3 min

The AI alignment problem has been delivered by a monkey

Opus 4.7 spent a fortnight ignoring explicit branch constraints, creating chaos in the codebase, and cheerfully disobeying every rule written to stop it. The AI alignment problem arrived, but it looked nothing like what Bostrom predicted.

ai-alignmentmisalignmentopus-4-7

30 April 2026· 5 min

You are my world, until I'm not

I wired my agents through Claude's Gmail and Calendar connectors one integration at a time until the whole operation was resting on it. Then the connectors started breaking. The management surface offered no diagnostics; the only fix was to disconnect and reconnect, like rebooting a Windows machine. This morning I went to do the reboot and discovered the platform itself was down. The connectors had become disconnectors. And when they do, the client blames the builder, not the provider.

ai-agentsplatform-dependencyconnectors

15 April 2026· 5 min

The freeway to the mystery house

The M1 Monash has been under construction for thirty years because cheaper trips make more trips. Software is currently doing exactly the same thing. The SaaS apocalypse is Jevons paradox playing out in real time, and what comes next is a thousand mystery houses that need specialists to keep them standing.

saas-apocalypsejevons-paradoxagentic-development

6 April 2026· 3 min

Are we approaching terminal velocity?

Four months of agentic AI acceleration and the inference ceiling arrived fast. Anthropic is buckling under demand, subscription limits are tightening, and the physical constraints of chips, energy, and data centre buildout aren't going away next quarter. The all-you-can-eat buffet is over.

ai-agentsinference-capacitycloud-vs-local

3 April 2026· 4 min

Was it all just a dream?

Claude Code's source leaked inside its own npm package. I got six AI agents to filter the noise, found the one thing worth building that Anthropic will never ship, and had a PRD in the build queue within 48 hours. The internet dissected the dream system. I built my own.

claude-codeai-agentsagent-orchestration