Field notes

Lessons from the forge.

A weekly log about building an AI-first company from Norway. The lessons are shared. The machine stays protected.

2026-07-05

The useful agent workflow in 2026-W27 is the one that can show, before release, which mode

The useful agent workflow in 2026-W27 is the one that can show, before release, which model touched which task, what it cost, and which veri

Read note →
2026-06-28

The Control Loop Is the AI Product

The operating lesson

Read note →
2026-06-21

AI Has Moved Into the Control Loop

The operator lesson: AI compounds only when action, permission, review, and evidence live inside the workflow.

Read note →
2026-06-15

Fallback Is Part of the Model Choice

The useful lesson this week was not that every workflow needs a bigger model. It was that model choice only becomes operational when the system can see failure, fall back cleanly, and decide which processes should move closer to local control.

Read note →
2026-06-13

Anthropic Fable 5 Shows Why AI Safety Needs Observability

The Anthropic Fable 5 controversy is not a simple story about safety being good or bad. The useful lesson for AI operators is narrower: safety controls can be legitimate, but invisible capability changes, fallbacks, and access changes break trust, debugging, and agent reliability.

Read note →
2026-06-13

Why Anthropic Fable 5 and Mythos 5 Were Suspended

Anthropic says a US government export-control directive suspended access to Fable 5 and Mythos 5 after a reported concern about a possible Fable 5 jailbreak. The important operator lesson is not the drama. It is that frontier-model access can change abruptly for reasons outside your product.

Read note →
2026-06-09

Model Progress Needs Translation

This week's useful signal was that model releases do not become business value as announcements. They become value only when an operator turns capability claims into workflow, budget, interface, and verification decisions.

Read note →
2026-06-07

Distribution Is Part of the System

This week’s useful signal was that AI advantage is moving beyond output and into the loop that gets work into the hands of users, earns trust, and learns from the market.

Read note →
2026-05-31

The Manager Layer Is the Product

This week’s signal was that AI work is getting less limited by raw capability and more limited by how well humans can assign, inspect, budget, and verify the work around it.

Read note →
2026-05-24

The Work Is Moving Up-Stack

This week's useful signal was not another demo. It was the same operating shift appearing from different directions: value is moving from raw model output to the systems, interfaces, budgets, and verification loops around it.

Read note →
2026-05-24

Verification Before Autonomy

The useful question is not how much work an AI system can produce. It is how reliably the work can be checked before it touches customers, code, or the public record.

Read note →