Atlassian's human-owner rule for AI agents is also how the seat bill survives
Atlassian co-founder Mike Cannon-Brookes wants a named human accountable for every AI agent action, and the Agentic Multiplayer Protocol shipped October 7 stamps that name into the audit trail. The company's own conference data say agents raise output while finished work does not rise significantly, because human review becomes the bottleneck; the November 5 report will show whether metered agent work adds revenue beside the seat.
Richard Tang · 4 min read
Every agent action will carry a human's name
Atlassian+2.23% — Atlassian, up 2.23 percent today co-founder and chief executive Mike Cannon-Brookes has one rule for workplace AI: whoever orders the agent owns what it does. "It has to be with the end user, ultimately," he said this week 1, and he told Fortune he does not want "the AI did it" to become an office excuse 2. The rule is sound governance. It is also how a company that bills by the seat lets agents in without unbundling its own revenue: a named human on every action is a human in a seat.
The audit trail became a product on October 7
The plumbing arrived first. At its Team '26 Europe conference in Amsterdam, Atlassian announced the Agentic Multiplayer Protocol (AMP) on October 7: every agent gets its own identity and a clear owner, runs under a named user's permissions or a dedicated service account, and writes every action to a full audit trail 3.
The logs do more than record that an agent acted; they record that an agent invoked by a named person acted 4. Cannon-Brookes treats attribution as a design problem: when a coding agent opens a pull request, a bundle of proposed code changes submitted for review, the question is whether the developer did it, the agent did it, or the agent did it on the developer's behalf 5.
Agent output now queues behind human review
His own conference data explain why the owner matters. At software firms that adopted coding agents, code volume and pull requests rose sharply while finished work did not rise significantly, and time from pull request to merge, the wait for a reviewer's approval, climbed by half as review became the bottleneck 4. Agents did not take the human out of software delivery; they moved the human to the end of the line, where every unit of machine output waits for a named person's judgment. Directing agents and reviewing their output is the skill worth building now.
Coding agents raised output and review time; finished work did not rise significantly
Data
| Value | |
|---|---|
| Lines of code | +30% |
| Pull requests | +23% |
| Time from pull request to merge | +49% |
The seat survives, with a meter beside it
That queue is also the business model. Atlassian, the company behind Jira and Confluence, collects $6.5 billion in annual recurring revenue 5, and its customers are moving from licensed server products to cloud subscriptions for those tools 6. A rule that keeps a named human on every action keeps humans in seats, whatever the intent.
A second meter is being wired in beside it. AutoDev, an agent now in early access, takes small bug fixes and feature requests on an assigned credit budget, so that work is priced by the unit 4. ServiceNow+1.70% — ServiceNow, up 1.70 percent today is bidding for the same ground: its AI Passport, unveiled on October 15, checks each agent action against its permissions at runtime and is sold as seats plus AI credits 7.
Each agent action still needs a person who reviews it, and that person holds an Atlassian seat. Reviewing agent output is the work Atlassian sells.
November 5 prices the doctrine
The buy side has already paid for the story. Atlassian's June quarter brought $1.77 billion of revenue against $1.66 billion expected, and revenue over the trailing year reached $6.6 billion, up 26% 8. Mizuho lifted its target to $235 this month on cloud demand and AI interest 8; Wells Fargo−1.93% — Wells Fargo, down 1.93 percent today sits at $230 on early adoption of Rovo, Atlassian's AI assistant 6. The stock touched its annual high of $223.70 while retail traders on Stocktwits, a social trading chat, stayed bearish 6.
Revenue has grown at least 14% year over year for twelve straight quarters
Data
| Revenue | |
|---|---|
| Q3 '23 | $0.98B |
| Q4 '23 | $1.06B |
| Q1 '24 | $1.19B |
| Q2 '24 | $1.13B |
| Q3 '24 | $1.19B |
| Q4 '24 | $1.29B |
| Q1 '25 | $1.36B |
| Q2 '25 | $1.38B |
| Q3 '25 | $1.43B |
| Q4 '25 | $1.59B |
| Q1 '26 | $1.79B |
| Q2 '26 | $1.77B |
The weak link was named the day AMP shipped. A record that someone clicked approve does not show meaningful human review, and the identity controls that would capture real judgment are the piece Atlassian labels "soon" 9. Whoever builds that layer, Atlassian or ServiceNow, owns the rails that agents run on.
The lines that would prove a second engine
The fiscal first-quarter report lands November 5 after the close, the first print that can show whether credits and Rovo attach become revenue 8. The lines to watch are any credit or consumption revenue beside seat and cloud growth: if metered agent work shows up as revenue, the doctrine has a second engine; if the bill still moves only with seats, the agents are a feature that sells seats.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
