The reasoning, in full.
Long-form pieces that teach a method or argue a thesis, with the evidence shown rather than asserted. These hold up regardless of when you find them — there is no news cycle to catch up on first. Looking for something shorter and more current instead? See What We Think.
- Understand · 6 September 2026
The bottleneck moved
For most of AI's useful history, capability was the limit on what it could do for a business. That's closing — and the bottleneck left standing is whether anyone has ever said, precisely, what the job actually is.
- Understand · 5 September 2026
The phase you're stuck in without knowing it
Adoption of AI is high and climbing; the financial return from it mostly isn't. The three stages — access, harness, role — that explain the gap, with almost all the return sitting in the third one.
- Understand · 4 September 2026
The signals nobody's taught you to read
The specific, learnable signals that tell you a piece of your business is ready for an AI employee — usually a $2,000-$8,000 consulting conversation, explained here directly.
- Understand · 3 September 2026
Why "off the shelf" fails here
A generic agent-builder platform gives you somewhere to configure a role. It doesn't define the role for you — the same reason buying accounting software doesn't create a CFO.
- Understand · 2 September 2026
What it looks like to actually declare a role
A concrete walkthrough of what it actually takes to declare one real piece of a job precisely enough for an AI to operate it: the owner, the boundary, the material, and the check.
- Understand · 1 September 2026
Governance is the seatbelt, not the destination
A written boundary and an enforced boundary are two different things. Why holding one requires something outside the model entirely, and what that already looks like, built and free.
- Sentinel · 29 August 2026
Your scanner and npm may not be reading the same package
A tar archive can hold two entries at one path, and the tools that read one don't agree about which wins. An adversarial review found our own tool gave three different answers to that question — and why the fix was to refuse the input rather than pick a side.
- Sentinel · 17 August 2026
The package you approved isn’t the one running tomorrow
Why a routine MCP server release is the moment your agent’s behaviour changes without anyone deciding it should — what Sentinel records, the coverage figure we retired for being unreproducible, and the flaw an adversarial review found in our own parser.