Atlassian Team '26 Europe: move to AI as team sport with 'multi-player' agentic system

You have probably watched a colleague lean into a chat window, ask an agent to draft a brief or summarise a thread, then copy the result into Slack or a Google Doc. That workflow treats AI as a vending machine: prompt in, artefact out, human ferries it to the next stop. Atlassian's new Agentic Multiplayer Protocol (AMP) targets that handoff. If agents stay locked in one-to-one sessions, teams still work in series and nobody else sees what the agent did or why.

Mike Cannon-Brookes, Atlassian co-founder and CEO, framed the shift at the company's European customer conference in Amsterdam. "The agent experience is largely single player today," he told a press briefing the evening before the keynote. "It is largely on my computer, I talk to an agent, maybe it's running in the cloud or running locally, and it's me back to the agent and I get some output and I take the output somewhere else." In a statement he added that "the companies that pull ahead will be the ones that get their people and agents working together in the flow, out in the open, as one team", and that AMP is how Atlassian is making that real.

What AMP does

Sherif Mansour, head of AI at Atlassian, says most people picture an AI experience as a chat prompt with an agent. That suits individual productivity, but it breaks down when you are working on a document with another teammate and an agent at the same time. Atlassian builds apps for collaborative teamwork, usually bridging technology and business teams, so it mostly makes multiplayer products, and AMP brings multiple specialist agents into that model.

AMP lets agents join work where it already happens. On the Atlassian platform, people bring agents in through @mentions in Confluence, Jira comment threads or Loom video briefs, while the Atlassian MCP (model context protocol) server pulls in context from third-party tools such as Figma or IDEs. Each agent has a clear owner and a distinct profile, and agents appear in real-time presence bars and cursors alongside their human teammates, so the whole team can see an agent's work and the context around it. Atlassian grounds the agents in its Teamwork Graph, which connects more than 250 billion objects and relationships and, through a new Code Search app, now reads source code down to the function, symbol and class level.

Atlassian also wants to drag hidden agent work into the open. As AI adoption grows, the company argues, more work happens where leaders cannot see it: in terminals, local sessions and third-party bots. A feature called agent sessions surfaces cloud and local agent work in Jira and the Teamwork Graph, so context carries forward and the work stays visible, traceable and governable.

Open to other vendors' agents

Cannon-Brookes pitched AMP as a hedge against lock-in to any single frontier model. "Most customers, well north of 75% of our customers, are using multiple large-scale vendors," he said. "They're picking multiple foundation model vendors, often with Microsoft in the mix, often with Google in the mix." Mansour says interoperability runs both ways: other app vendors can work with the Atlassian platform, and Atlassian's agents can work inside theirs. "It's huge part of AMP to let customers pull in the agents they want. It doesn't have to be ours," he says. Atlassian's own AI platform, Rovo, launched in October 2024.

The coordination problem

Dave Meyer, head of product at Atlassian, put the business case in blunter terms. Models keep improving, he said, "and yet we are not seeing super revenue at every company that's building software". Teams are building more software, but coordination between them is getting harder. "We're seeing more quality problems, more review problems, more rework because we have to re-implement and re-architect systems multiple times because it wasn't fully built or fully planned out in our enterprise architecture to begin with," Meyer said. He believes those large-scale problems sit on top of a deeper one: "we're only scratching the surface of the more low-level knowledge and context loss that's happening".

What this means for L&D and innovation leaders

The shift from solo-agent prompting to multiplayer agentic work changes the capability map for learning and development. When AI lives in a chat sidebar, training focuses on prompt craft: how to write a clear instruction, how to iterate on a response, how to fact-check the output. Those skills remain necessary, and they stop being enough once agents join a live document or project board. Teams need to learn how to set task boundaries with an agent that sees the same page, how to resolve conflicting edits and which decisions stay human.

Named owners and visible agent sessions also give governance something to hold on to. If an agent edits a Confluence page or a Jira ticket while two people are working on it, the team needs a common language for who approved which change and a process for rolling back agent contributions that miss the mark. Meyer's point about rework and context loss is the practical stake. Faster code generation only pays off if teams can coordinate the output, and that is a team skill, so the next wave of AI capability building should spend less time on individual productivity tricks and more on the team protocols that let people and agents work on the same thing at once.

Sources:

• Computer Weekly, 7 October 2026