Why “AI coworker” is not a marketing word
Chatbots answer questions. Coworkers ship work. The line between them isn’t branding — it’s memory, action, and accountability.
For most of 2024 we tried to call Octodus an "AI assistant." It tested fine. It also lied about what we built.
An assistant is the person who hands you the pen and waits. They schedule. They forward. They produce a draft for you to fix. They never ship without being told to, and they never decide whether to ship at all. That’s a useful job — humans do it well, software has done it for thirty years — but it’s not what teams kept reporting back to us.
Octodus, in those same threads, was the one who decided the pageant banner went to print, which 47 of 110 prompts shipped to the dev team, and that the disk cleanup ran at 3:14 AM without waking anyone up. Those aren’t assistant moves. Those are coworker moves.
Three things a coworker does that an assistant doesn’t
- Remembers the team’s rules.Not a system prompt — actual rules learned from being told once. "Draft, wait for OK, then publish." We never get asked again.
- Takes action, not notes.A coworker files the task themselves. They run the script. They post the receipts. They don’t hand you the script in a chat bubble.
- Owns the outcome.If the cleanup ran but the disk is still full, that’s on the coworker to figure out. Octodus posts the before-and-after — and re-runs if it has to.
A coworker is judged by the work that got shipped — not the work that got described.
Why the language matters
It changes how you deploy it. If Octodus is an "assistant," you ask it for a draft, you read the draft, you copy-paste the draft, you ship the draft. If Octodus is a coworker, you say "this is going out at 17:00 — make it happen" and you go to lunch. The product is the same — the contract is different.
It also changes the bar. Coworkers can be wrong. Coworkers can be told off. Coworkers don’t get a free pass for hallucinating because they’re "just AI." If we ship a "coworker," we are accepting a higher accountability than we would as an assistant. That’s on purpose.
What this means for what we build next
Two things. First: memory and rules get more investment, not less — that’s the substrate that lets a coworker stay accountable across days, not just messages. Second: every new capability we ship is now scored against a single question — "could a human teammate own this end to end?" If yes, Octodus will too. If no, we don’t ship the capability yet.
We’ll see how the language ages. But after a year of running this experiment in production threads with several thousand teams, "AI coworker" has stayed honest — and that’s a higher bar than most marketing language clears.
