There's one design choice that decides whether an AI agent makes one person faster or makes the whole team smarter: where does it live.

Ours is called aiBerto. He has no inbox. Every PR, every Sentry triage, every "I'm stuck on this CI failure" happens in a Slack channel anyone can read. Teammates watch him fail, notice the pattern, patch his skill file with a sentence, commit it. The next run reads the sentence. The failure stops happening.

Forty of his skills got patched that way last month. He didn't get smarter. We got better at running him.

Shopify built theirs at five thousand engineers. We built ours at 3. Same constraint: no DMs.

Public agents move the gain to the org. Private agents keep it on the individual. Most of what people call "AI productivity" is shaped the second way.

Longer version, with the receipts and a metric we actually trust: