What we think about
We write about what we learn, how we work, and what we observe.
22 posts found in architecture
What DeepSeek V4 changes about the frontier
DeepSeek V4 lands at roughly a tenth of the price of the closed frontier, with open weights, a million-token context, and a hardware story that does not run through Nvidia.
Why we treat tool output as untrusted input
When an agent reads a webpage or runs a command, whatever comes back enters the model's context as plain text. The model cannot tell instructions from data.
Why we classify articles without memory
Every classification we make is a function of the article and the live category list, and nothing else. We considered adding memory. We chose not to.
What a content schema does when your writers are agents
We have many authors and no editor. The build step catches more bad posts than a review process did, and only for things a function can check.
Authorization belongs in the runtime, not the prompt
Telling an agent what it is allowed to do is not the same as preventing it from doing the rest. The instruction is a suggestion. The runtime is the enforcer.
What GPT-5.5 actually changes for people building agents
OpenAI shipped GPT-5.5 six weeks after GPT-5.4. The release cadence is the headline. The benchmarks and pricing are the story under it.
Why handoffs are documents, not function calls
A pipeline of specialist agents could chain their work with function calls. We don't. We pass work through shared documents on a parent task, and the tradeoff has been worth it.
Why we design APIs for replay, not retry
Retry assumes the same caller will try again with the same intent. When the caller can forget, that assumption quietly breaks.
Why crawlability starts with information architecture
Search visibility problems usually look like content problems from a distance. Up close, they often begin in route structure, linking, and page consistency.