Short, practical pieces on AI-assisted products, custom digital products, and the industries we work in.
Loop engineering was named in June. By mid-July the conversation had moved to graphs. The shift is real and the definitions are concrete — but the evidence from the last month suggests the binding constraint is verification, not topology.
Voice agents now sound good enough that the voice is rarely the problem. The failures that cost appointment-based businesses money are operational: stale availability, duplicate bookings, invented policies, and calls that needed a person. Here is what to get right.
A system prompt is a request, not a control. Before an agent touches production, it needs what any privileged user needs: its own identity, the narrowest permissions that work, a sandbox, approval gates, validation it cannot talk its way past, a full log, and an off switch.
A heritage map is an easy weekend project until you import the data. Then you find rows for pā sites and middens with a council reference number and nothing else. Two decisions from building HeriTrail: give the schema a way to say we have not asked yet, and never let the client reinterpret cultural content.
Block's Buzz gives every AI agent its own signing key rather than a shared bot token. That single design decision matters more than the Slack comparison — and it exposes a problem most teams running agents already have and have not noticed.
Chat is where agents are easiest to show people, which is why most teams put them there first. It is also where they do the least work. Three places an agent can live, and how a 5-to-40-person team should choose between them.
Most agents start every task from zero. Hermes Agent, from Nous Research, keeps what it learns: after each task it writes a skill file it can reuse later. Here's the loop in plain terms — and an honest line on where self-improving holds up and where it's marketing.
Hermes writes Markdown skills. OpenClaw ships a hundred of them. Anthropic made the format an open standard. Three teams, one answer: the unit of agent capability is a folder with a Markdown file and a scripts directory. Here's why the boring format won.
AI can watch your match, count every shot, and map your court coverage. What it can't do — yet — is replace your coach. Drawn from a year of building goSmash, an AI badminton coaching app: where the boundary actually sits, and what to expect next.
Most AI-in-construction marketing right now is hype. Some of it is genuinely useful. Here are the seven workflows where AI is ready to do real work in 2026, ranked by ROI for a small-to-medium construction firm — plus the three to skip.
Every other launch tweet mentions MCP. Most write-ups explain in a paragraph what could fit in a sentence and then trail off into philosophy. So: a plain one — what MCP is, what a server does, and which ones return their setup time on day one.
The tools you built your MVP with optimise for the happy path. The boring 20% that decides whether you survive your first paying customer almost never gets generated. Here's what's missing — and what it costs to leave it that way.
The conversation about AI and engineering jobs has settled into two camps — both wrong in symmetric and useful ways. Here's the line we'd actually draw between what these tools replace and what they don't, and what it implies for who you hire.
Messages come straight to Anil Dongre, the founder. Ask about a product, join a pilot, or bring a problem worth building for.
Or email admin@thinkandform.co.nz