Three lessons from building the AI assistant: train one skill at a time, adapt the workflow to the agent, and build a continuous improvement loop.

01

One skill at a time

Building a coworker is like hiring a human — no throwing it in the deep end. You garden it: get one capability solid, then add the next.

02

Adapt the workflow to the agent

The agent often has its own way of doing things. Find the workflow that's optimal for it, then teach the humans to work that way.

03

Build a continuous improvement loop

Most of our compounding wins came from a second agent, Codex, running loops to watch how Ace performs and steadily improve it.