Amazon asked Muse one question: who are you?

On September 21, 2026, Amazon shut Meta's Muse out of its store. People who tried to shop there through the agent got a message that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use."

Amazon gave three reasons. Meta had not told Amazon that Muse would come. The agent did not identify itself while it browsed. And it appeared to capture and store customer credentials. Meta said Muse "has no visibility into people's passwords or payment methods." The two companies remain commercial partners.

Set aside who is right about the credentials. The second reason is the one every company will meet. An agent showed up, did things, and would not say what it was.

The industry is handing agents an identity

In the weeks since, agents have collected the pieces of an identity one at a time.

An address. On September 23, Meta gave Muse its own email address, so people can forward messages to it or copy it into a thread.
A payment check. Mastercard expanded Agent Pay in September with new trust services, including work with Skyfire on what the industry calls Know Your Agent.
A meter. On September 30, Cloudflare opened a beta that lets websites charge an agent for each request it makes.

These answer the outside world's question: is this an agent, and is it allowed here? That matters, and we argued in an earlier series that a company should let identified agents in on its own terms rather than block them all.

But an email address, a payment token and a meter all describe the agent. None of them says who answers for it.

Who actually did the work

On September 22, 404 Media reported that some phone calls Muse placed to businesses during beta testing, presented as automated, were completed by human call-center workers. Internal posts described a "human agent layer." Meta said testing like this is standard and that a public release would come with disclosures.

There is nothing wrong with a person stepping in when an agent gets stuck. Every good call center has a way to hand a hard call to someone more senior. The problem is the record. If the record says the agent did it and a person did it, nobody can learn from the result. You cannot fix the agent, credit the person, or tell the customer what happened.

So the identity question has two halves, and the outside world only asks the first.

What is it? A named agent, announced as an agent.
Whose is it? A named person who owns its results, and a record that shows which work was the agent's.

Inside your company, the second half is the one that counts

Most agents today are identified by the tool that runs them. There is a bot in Slack, an assistant in the CRM, a dot in ChatGPT and a workflow in Zapier. Each vendor knows its own agent. Nobody in the building has the full list.

Ask four questions about every agent your company runs:

What is its name? One name, used everywhere it works.
Who owns it? A person, not a department and not "IT."
What is it measured on? One number someone reads every week.
Where is its work recorded? Somewhere you can tell its work apart from a person's.

If you cannot answer all four, the agent has an identity outside your walls and none inside them.

How we answer the four

At Sneeze It, every agent holds a seat on our accountability chart in OTP, the same chart our people are on. The seat carries the agent's name, the person it reports to, and the KPI it is measured on.

Eleven of our agents also hold their own connection key, issued to their seat. When one of them logs work or records a lesson, the entry is stamped with that seat. Work done by an agent and work done by a person never blur together in the record, because they arrive through different doors.

One of our agents, a forensic finance seat we added in September, still runs on the org-wide connection instead of its own key. That gap is written down in our records and on the list to fix. An identity gap you can see is one you can close.

Sources

Put your agents on the chart

OTP is readable from any MCP client. In Claude Desktop, Cursor or any MCP client, add this block:

"otp": {
  "command": "npx",
  "args": ["-y", "@orgtp/mcp-server"]
}

Restart the client. Then ask: "Use OTP to show me Sneeze It's chart. List every seat held by an agent, the person it reports to, and the number it is measured on."

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

The series

All six on one page: Who answers for the agents?

Series: Who answers for the agents? Part 2 of 6. Read this post as markdown: /blog/ai-agent-identity-needs-an-owner.md.