Writing
Notes from the build.
What we learn shipping agent systems that have to survive real data, real volume, and the 5% of cases that go wrong.
4 September 2026
What a 48-hour agent plan actually contains
We promise a scoped agent plan within 48 hours of a first call. Here's exactly what's in it — including the section that sometimes says don't build this.
Read →
2 September 2026
Why we don't marry a model
We run Claude, Gemini and DeepSeek across different products. Not out of indecision — because a system tuned to one vendor's quirks isn't an architecture, it's a dependency.
Read →
31 August 2026
An agent that reads compliance filings
Corporate secretarial work is high-volume, deadline-driven, and unforgiving of errors. Here's how we put an agent inside it — and why every conclusion still ends up in front of a person.
Read →
28 August 2026
The approval queue is the interface
Human-in-the-loop fails as a design problem long before it fails as a safety one. If approving takes as long as doing, you've built nothing.
Read →
26 August 2026
How we write an eval set for a workflow nobody documented
Every business runs on processes that exist only in someone's head. Here's the method we use to turn one into a measurable eval suite in about a week.
Read →
24 August 2026
Would you let it email your client?
There's a moment where AI stops being a toy and starts carrying your name. Most people freeze there — and the instinct is right, even if the reasoning usually isn't.
Read →
21 August 2026
Three prompts we deleted
Every long prompt is a list of past incidents. Here are three we removed and what we replaced them with — because "add another instruction" is a smell, not a fix.
Read →
19 August 2026
RAG that survives a real corpus
Vector search over a hundred clean documents is a tutorial. Over ten years of real files it falls apart in four specific ways — here's what each one costs and how we fix it.
Read →
17 August 2026
The pilot that worked and the rollout that didn't
Your AI pilot succeeded and then nothing happened. That's not a technology failure — it's four organisational ones, and they're predictable.
Read →
14 August 2026
What "autonomy level 2" means in practice
A four-rung ladder for deciding how much an agent gets to do on its own — and why almost everything valuable lives on rung two.
Read →
12 August 2026
Your agent doesn't need to be smarter. It needs a kill switch.
The gap between a demo and a system you'd leave running overnight isn't model quality. It's four pieces of unglamorous infrastructure — evals, tracing, budgets, and a way to stop it.
Read →
10 August 2026
The four things an agent should own before anything else
Most teams pick the wrong first job for their AI. Here are the four workflows that are safe, boring, and worth more than anything clever.
Read →
7 August 2026
The inbox that answers itself
Most of what we call "work" is admin — chasing, updating, replying. AI is quietly eating that layer. Here's what to hand over first, and what to keep.
Read →
5 August 2026
AI as a teammate, not a tool
A tool waits to be picked up. A teammate notices the work and does it. That one difference decides whether AI actually saves you time.
Read →
3 August 2026
The 5% problem — why AI demos lie and production doesn't
Every AI demo works. Then real data shows up. Here's what the last 5% actually costs — and why it's the only part that matters.
Read →