blog
Writing
Notes on durable execution, silent partial failure, and recovering the one item that broke — in long-running AI pipelines and agent jobs.
Your agent isn't slow because of the CPU
A twenty-turn agent spends almost all its time waiting, and no hardware changes that. What does: never running turn 1 again because turn 18 failed.
What it means to be fault tolerant in the AI age
Notion pulled every Anthropic model from its picker — not because the API was down, but because it was degraded, returning 200s that didn't work. That's the failure nothing in your stack was built to catch. Here's the layer that catches it.
Harvey built our pitch deck. They were writing about themselves.
One of the most valuable AI companies just explained why it runs its own agent runtime. Multi-model, zero retention, cost — three reasons that aren't about law at all. They're the case for the workload layer.
Running an AI agent in production is six services in a trench coat
From the outside it's one product. From the inside it's twelve vendors holding hands. Here's the layer that's actually missing.