Writing
Notes from production
Incidents, post-mortems and design arguments from running autonomous agents that spend real money. Mostly about the failures that don't look like failures.
-
Three shapes of a runaway
Five months of production agent logs, and almost every expensive incident is one of three patterns. None of them look like failures while they're happening.
-
Prompt injection is a blast radius problem
You cannot reliably detect the attack. So stop making detection the control, and start bounding what a successful one costs you — including being honest about which bounds you actually have.
-
Reading usage without reading prompts
A guardrail sees everything an agent says. Here's how to build one that enforces a spend cap without ever being able to reconstruct a conversation.