- 2026-08-21
Govern the architecture, not the model
Trust is not the variable. A machine-checkable governance layer for agent fleets, and an honest account of how little it has been tested.
- 2026-08-21
AI literacy is now a civic question in Wellington
Thousands of public sector roles are going in the region. What the Impact Lab model offers a capital city that has to re-skill fast.
- 2026-08-19
I made my site agent-ready in an afternoon. Half the checklist was theatre.
Markdown for agents, content signals, and llms.txt are worth doing. Fake agent cards are not. A practical walkthrough with the receipts.
- 2026-08-17
Match the autonomy to the blast radius
A practical method for deciding how much autonomy an AI agent should get: start from the worst thing its permissions allow.
- 2026-08-15
The agent that improves itself through pull requests
A self-improving agent is a governance problem until the improvement loop runs through version control. The pattern in detail.
- 2026-08-13
Ask the partner for their problems: lessons from a civic hackathon
What running New Zealand's first Claude Impact Lab taught me about problem statements, team drafting, and judging on usefulness.
- 2026-08-11
An emergency tool that needs the internet is not an emergency tool
The Wellington Claude Impact Lab's winning team ran everything on local models. The design principle generalises to all resilience software.
- 2026-08-10
Why technical leaders should stay close to implementation
The moment a technical leader stops understanding the system at implementation level, their strategy starts to drift from reality.
- 2026-08-10
Governance is an enabler for agentic systems
Controls are not the brake on agentic AI. They are the reason you can say yes to autonomy at all, written by someone who builds both.
- 2026-08-10
Production AI agents need identities, not shared credentials
Why an AI agent doing real work needs its own scoped identity, and the credential patterns that make autonomous systems defensible.