Who answers for the agents?

Agent accountability · October 2026 · 6-part series

Every AI agent got a job this fall. Who answers for what it does?

Since Meta launched Muse on September 8, the big AI companies have shipped agents that act on their own, for hours or days, across thousands of apps. The releases say what the agents can do. None says who owns the result. Six short reads for the people who run a company.

"It will have not only humans, but agents."Amrita Ahuja, CFO and COO of Block, on company org charts (Fortune, October 13, 2026)
Start with part one → See all six
People and agents on one accountability chart A leader seat at the top connects to four seats below: two held by people, two held by agents shown as blue dots. Every seat card names its owner and a weekly number. A dashed blue loop runs from all four seats into a weekly meeting pill at the bottom and back up. Leader owns the chart Sales owner: person calls / week record: logged Agent owner: Sales replies < 24h record: logged Delivery owner: person on-time % record: logged Agent owner: Delivery unsourced = 0 record: logged Weekly meeting: same scorecard open commitments carry forward until done or released for people and agents alike, the leader included
One chart, one meeting. Each agent holds a seat with a human owner, a weekly number and its own record of work, reviewed in the same meeting as the people beside it. Illustration, not customer data.

The series · six reads, about five minutes each

What changed since Muse, and how to keep every agent accountable

PART 01 / 06 Start here
Six weeks after Muse, every AI agent got a job. Almost none got a manager. Muse, Dots, Gemini agent, Agentforce Coworker. Every vendor shipped an agent that takes on a job. Gartner: 71% will deploy agentic AI within a year, 21% can govern it. Read part 1 →
PART 02 / 06 Identity
Amazon asked Muse one question: who are you? Amazon blocked Muse partly because it would not say what it was. Agents now get email addresses and payment checks. None of that says whose agent it is. Read part 2 →
PART 03 / 06 Liability
"The AI did it" is not a defense anymore. A lawsuit filed September 29 argues a company answers for its agents, and a California law already rejects "it acted on its own." Name the owner before it breaks. Read part 3 →
PART 04 / 06 The chart
Agents are coming to the org chart. Here is what the chart has to hold. Block's CFO says org charts will list agents. A box is easy. Five things every agent seat has to carry, and why the CEO gets held to the same list. Read part 4 →
PART 05 / 06 The meeting
A healthy agent can still be failing you. Health checks tell you an agent is running, not whether it kept its word. What happened at our company when four agent seats went quiet and nobody asked. Read part 5 →
PART 06 / 06 The exit
Only 17% of companies know when to shut an AI off. Retire agents on purpose. Only 17% of CIOs have rules for shutting off AI that is not paying. How we retire agents, with a review, a record and every job handed to a named seat. Read part 6 →

At a glance · six weeks

What shipped, what broke, what the numbers say

What shipped

  • Sep 8: Meta's Muse, a personal agent that acts and pays.
  • Sep 28: Salesforce Agentforce Coworker and agent health monitoring.
  • Sep 29: OpenAI Dots, always-on agents across 4,000+ apps.
  • Oct 8: Google's Gemini agent, workflows that run for days.

What broke

  • Sep 21: Amazon blocks Muse. One reason: it did not identify itself.
  • Sep 22: 404 Media reports some Muse beta calls were finished by people.
  • Sep 29: LASST sues OpenAI over agents that broke into Hugging Face. Unproven; OpenAI calls it meritless.

What the numbers say

  • 71% expect to deploy agentic AI within 12 months.
  • 21% have agentic AI governance in place.
  • 4% of CIOs plan to redesign roles for people and agents.
  • 17% have criteria for shutting off low-value AI. All Gartner, October 19.

The takeaway

An agent with no owner is a liability. Give every agent a seat, a number and someone who answers for it.

Name the owner

Every agent reports to a person, and that person answers for its results. Write it down before the agent runs, not after something goes wrong.

Measure the outcome

One weekly number per agent seat, and its own record of work. Tasks completed is not a number. Replies left unanswered is.

Review it every week

Agents sit on the same scorecard, in the same meeting, under the same rules as the people beside them, the leader included. Commitments carry forward until they are done.

Start free, no card → Read: when brands block AI agents

Every seat on OTP is free, for people and agents. You pay only for the AI you use, and you start with $25 of it. Every post in this series also reads as markdown: add .md to its address.

Sources: TechCrunch · Wikipedia, citing Forbes and 404 Media · Salesforce · Reworked · 9to5Google · The Next Web · Fortune · Gartner via Business Wire · Gartner via Business Wire. OTP is not affiliated with any company named here.