Blog

Notes from the runtime.

How we think about running agents in production, and what we're learning building it.

Verification is the hard half of agents

Everyone can generate an agent. Almost nobody can tell you, at 3am, whether it did the right thing. That gap is the whole product.

Checkpointing sounds simple until a tool call is non-idempotent. Here is how we make a paused run resume exactly once.

"It felt better" is not a deploy signal. Score every prompt change against a suite and let the build tell you the truth.