When a personal agent marks it done, the record should say so.

Personal agents act through your logins. Claude Duet uses the apps you connected to Claude. Muse uses the services you connect one at a time. Dots use the apps you plug in. Anthropic is careful that Duet "never directly sees your passwords," but from the app's point of view, the request still arrives as you.

So when Duet updates a deal in your CRM at 2 a.m., the CRM says you did. When a dot closes a ticket, the ticket system says you did. The record is accurate about whose authority was used and silent about who actually did the work.

Why "who did it" matters

Three ordinary situations show why.

A number looks wrong. Your weekly sales number jumped. Did the sales lead update it after checking, or did her agent update it from an email it misread? You will read the number differently depending on the answer.

A commitment closed early. A to-do due Friday was marked done on Tuesday night. If a person did it, you trust it. If an agent did it, you might check.

Something went wrong. A customer got the wrong message. You need to know whether to coach a person or change an agent's permissions. Those are different fixes.

In each case you need to know whether a person or their agent did it, and which agent. Whose account was used is not enough.

Email already solved this

Executive assistants have sent email for their bosses for decades. Email clients handle it with one line: "sent by Alex on behalf of Jordan." Everyone who reads it knows two things at once: Jordan stands behind it, and Alex typed it.

Company records need the same line for agents:

Marked done by Claude Duet on behalf of Maria Lopez, Tuesday 11:42 p.m.

That line keeps Maria accountable, because it was her agent and her approval, and it tells everyone else how the work actually happened.

Three things the line has to carry

The agent. Duet, a dot, Muse, Gemini agent. Different agents behave differently. You want to be able to see that one of them is misreading threads more than the others.
The person. The human the agent works for, who stays accountable.
The approval, if one was needed. If the agent changed something someone else owned, who said yes.

None of this requires the agent vendors to change anything. The system receiving the change can record it, as long as it knows which client is calling. Most connections already announce the app they come from. Most systems throw that away.

The weekly question this answers

Once the line exists, a leadership team can ask a question it could not ask before: how much of last week's work did our agents do, and how much of it held up?

That is how a team starts managing personal agents with evidence. We wrote about why you cannot hold an agent accountable that you cannot see in Anthropic counted its agents. Can you count yours? The via line is how you count personal agents, which never show up as anything but the person.

Sources

Give your team's agents the plan

Personal agents read whatever they are connected to. OTP is where your company's chart, priorities, scorecard and commitments live, and any MCP client can connect to it. In Claude or ChatGPT, add a custom connector with this address:

https://orgtp.com/api/mcp

Then ask: "Using OTP, what are my priorities this quarter, what did I commit to in our last meeting, and what is due this week?"

Start free at orgtp.com. Every seat is free, for people and agents. You pay only for the AI you use.

The series

All ten on one page: Personal agents are coming to work

Series: Personal agents are coming to work. Part 6 of 10. Read this post as Markdown: /blog/agent-work-needs-a-via-line.md.