Field Notes from Running AI Agent Teams in Production
One studio, AI agent teams, real costs, honest failures. Most AI development content is vendor marketing or tutorials. This is what it actually looks like to ship products with agents. The agent workforce runs on Claude Code - Anthropic's CLI harness - with Claude as the foundation model.
New here?Start Here for curated reading paths, or readThe System for the operating model behind it.
Articles
Deep dives on a single topic.
The Grant Table Is the Kill Switch: Access Control for a Shared AI Employee
Several people share one autonomous agent: the identity provider proves who they are, but a grant read live on every request decides whether they still get in.
Read article →What an Agent May Do Is Not How It May Start
One authorization setting was answering two unrelated questions. Splitting exposure from initiation is what made the entitlement model enforceable.
Bake Every Author-Built Connector In, Keep It Inert Until Bound
When a vendor has no MCP server you author your own. The real decision is where that connector code lives: baked into one shared image, inert until bound.
Template Gives You Consistency, Not Credibility
Per-vertical content a practitioner respects cannot be stamped from one template. Depth at altitude takes serial authoring through an adversarial critique loop.
Ship Log
What changed this week.