Your AI Vendor Is Your Competitor
OpenAI and Anthropic raised $5.5B in a month to start consulting arms that compete with the firms deploying them. Here's how services firms survive it.
Blog
Notes on building the collaboration layer for agents and operators.
Two Agents Are Better Than One
Smarter models won't unlock the next productivity step. The unlock arrives when two agents act on the same state, and the providers can't host it.
Your AI Vendor Lock-In Is Five Layers Deep
Model portability is the easiest layer to solve. Enterprise AI lock-in runs five layers deep — and the deepest one cannot be exported.
The Collaboration Tax in Your Token Bill
Tokenmaxxing exposed the friction tax at the tool layer. The bigger line item is collaboration — agents and operators paying tokens to find each other.
How We Built Chief of Staff as a Mounted Agent
We built an agent that does not live on our servers. Three memory layers, two operators, one imprint. Here is the architecture and what surprised us.
What Opens Up When You Split the Agent From the Model
When you split the agent from the model, the economics flip: builders stop rationing intelligence, token efficiency becomes visible, and every model price drop passes through as a free upgrade.
The AI Agent That Survives the Session
AI agents should not live inside the runtime that executes them. Five things need to survive the session, and a five-question diagnostic to evaluate whether yours do.
The Ceiling on Cloud Agents
Cloud agents share a ceiling: capped capability, hidden drift, vendor-controlled lifespan. Three questions to ask any vendor before you bet on one.
The Spec Lived in Slack. It Didn't Survive the Weekend.
Why file-based spec distribution silently breaks multi-agent pipelines — and what a true shared source of truth actually requires.
Shared Agents, Not Shared Skills
Sharing a skill file between agents solved instruction drift. Scaling the pattern turned it into something else: shared agents loaded at runtime from published assets.