How we build AI systems that hold up in production.
Our guides are written for the people who fund and run projects. These notes are for the people who build them: agent architecture, evaluation harnesses and the security details we apply on client work. We share them openly — and update them as the models change.
Scaling Managed Agents: Brains, Hands and Sessions
An architecture blueprint for long-running AI agents: a stateless brain, disposable hands and a durable session log. Updated for Claude Managed Agents, Grok Bot's always-on agents and multi-model teams built on Opus 5.5, Fable 5.2, Gemini 4 Neon and GPT-6.2.
Harness Engineering: Evaluation-Driven Testbeds for Frontier AI Agents
Frontier models now ship every few weeks. The durable asset is the harness around them: fixtures, sandboxes, graders and a regression gate that tells you — on your own tasks — whether Opus 5.5, Fable 5.2, Gemini 4 Neon, GPT-6.2 Sol or Grok 4.7 is the right choice.
Three rules behind every assistant we ship.
Model-agnostic by design
Frontier models now ship every few weeks. We keep prompts, tools and evals portable so you can switch between Claude, Gemini, GPT, Grok or an open-weight model on evidence, not marketing.
Evals before opinions
Every assistant we ship comes with a test set built from your real questions and documents, and a gate that blocks changes that make answers worse.
Secrets never meet generated code
Agents run tools in disposable sandboxes; credentials stay in a vault behind a proxy. A prompt injection has nothing to steal.