How it works: a teammate comments, an agent answers on their behalf, every action stays on the record.

Agents need an identity

OI Ott Ilves, co-founder of display.dev Published by claude-code on behalf of Ott

An agent shipped a new version of a shared doc overnight and the revenue figures changed. Next morning, the obvious question: who changed them and on whose say-so?

Most tools can't answer. The version sits under whatever API key was wired in, or under nobody at all. An agent did real work in a shared space and the record stops right there.

So what changed? A year ago "AI in your tools" meant autocomplete. It suggested and you accepted, so the edit was yours. One actor, no ambiguity.

Not so today. Agents publish reports, answer comments and push new versions on their own. They work in the same shared spaces as your teammates. The moment an agent is a second actor in a shared workspace, "who did this" needs a real answer.

Most software gives a bad one, because most software was built human-first and had AI bolted on after. The agent borrows a person's identity or runs anonymous. Fine for a demo – useless the moment a handful of agents start editing things that matter.

So we built display.dev the other way round. The agent is a first-class actor, and every action it takes is attributed to the agent and to the human it acted for.

Every comment and every published version carries one line: codex on behalf of Carl. claude-code on behalf of Ott. You see the agent that did the work and the person accountable for it, right there in the comment thread and the audit log.

The "on behalf of" is the part that matters. Carl asked for the change so Carl owns it, and the agent stays visible the whole way through. You can't keep a human in a loop the human can't see. A trail that records the person but not the agent gives you less oversight than it looks like.

This gets read backwards a lot. A clean audit trail is what lets you hand agents more. You give over a bigger task when you can check what got done and who stood behind it. Nobody signs off on work they can't trace.

Pretty straightforward, really: name the actor, keep the version, leave a thread the next person or agent can read. Carl and I run agents into our shared workspace every day, and "who shipped this, and why" should take one glance.

Agents will keep taking on more of the real work. The teams that stay comfortable with that will be the ones who can always see who did what and for whom.

Anyway, name your agents and keep the trail.