Atlassian Launches AMP, Giving AI Agents an Identity, a Seat, and a Job on the Team

On October 7, at its Team '26 Europe conference in Amsterdam, Atlassian made a bet about where workplace AI goes next. The problem it named is easy to recognize: today's AI agents mostly work alone. An employee opens a chat window, gets some output, and carries that output somewhere else by hand. The agents are brilliant at individual tasks and nearly invisible to the team around them.

Atlassian's answer is AMP, the Agentic Multiplayer Protocol, a platform layer that lets AI agents participate in the same workspaces humans use, with a clear identity, a defined owner, and governance over what they can touch. Agents stop being hidden tools and start being visible participants: showing up in presence bars and cursors alongside human teammates, summoned by name in Jira tickets and Confluence documents, with their work traceable instead of vanishing into a terminal.

The framing comes straight from the top. Co-founder and CEO Mike Cannon-Brookes, in a statement accompanying the announcement, said: "The best work has never been a solo act, and now that's true of AI too. 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."

The best work has never been a solo act, and now that's true of AI too. 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.

The timing is deliberate. The announcement came a day after Atlassian said it was deepening its partnership with OpenAI to embed frontier models into its enterprise tools, and in the same week Cisco unveiled agentic plans for Webex. The enterprise software industry is converging, fast, on the same thesis: the agent era will be decided not by who has the cleverest chatbot, but by whose agents can actually be trusted to do work inside a company's real systems.

Every agent gets an identity

The heart of AMP is identity and visibility. Atlassian says that as AI adoption grows, more work happens where leaders and teammates cannot see it: in terminals, local sessions, and third-party bots. The company's response has two parts. Agent Sessions surfaces both cloud and local agent work in Jira and the Teamwork Graph, so context carries forward instead of getting lost in terminals. Non-Human Identities gives every agent, whether Atlassian's own Rovo or a third-party model, its own identity, so teams can see which agents are active, what each one can access, and who owns it.

This is the governance layer the enterprise AI market has been missing. A chatbot that reads your Jira tickets is convenient. An agent that can act on them is powerful, and slightly alarming. Atlassian is pitching scoped identities as the mechanism that lets companies hand agents real authority without handing them the keys to everything.

The pitch carries extra weight because of what it implies about lock-in. Cannon-Brookes told reporters that well north of 75 percent of Atlassian customers use multiple large language model vendors, presenting AMP as a way to avoid tying agent workflows to a single frontier model. Sherif Mansour, Atlassian's head of AI, said interoperability is central to the protocol: customers should be able to bring whichever agents they want into the platform, because the agents don't have to be Atlassian's.

In the flow, not in a tab

The second pillar is presence: agents show up wherever the work already happens. AMP lets agents participate through @mentions in Confluence, in Jira comment threads, or via Loom video briefs. A new Atlassian MCP server pulls context in from third-party tools like Figma and developer IDEs, so agents are grounded in the same materials humans use.

Record for Agent turns a quick screen-and-voice recording into an actionable brief that an agent can convert into Jira work items, a working prototype, or a Confluence specification. Interactive PR Reviews flips code review on its head: agents record a Loom walkthrough of their own changes in Bitbucket, explaining the key tradeoffs, so humans can judge the work in place.

Underneath all of it sits the Teamwork Graph, Atlassian's model of how a company's work fits together, which the company says now connects more than 250 billion objects and relationships. The graph is getting richer in this release: a new Code Search app reads source code down to the function, symbol, and class level; structured data becomes a native context type, so users can query data lakes and warehouses from anywhere through Atlassian Insights; and the library of 80-plus connectors keeps growing, with additions from Zoom, Gong, Microsoft Entra ID, and Google Identity. A new Artifacts feature gives agent outputs a permanent, governed home that indexes straight back into the graph.

Atlassian also says the new version of its platform runs leaner, using up to 25 percent fewer tokens for the same Jira and Confluence work in internal benchmarking on Claude models. In a world where agent workloads multiply token consumption, that efficiency claim will get as much attention from IT budgets as the features do.

The OpenAI layer and the real problem underneath

The AMP announcement did not arrive alone. On October 6, Atlassian and OpenAI said they were expanding their partnership to integrate OpenAI's frontier GPT-6 models into Atlassian's enterprise ecosystem, connecting the models to tools like Jira, Confluence, and Bitbucket, with the Teamwork Graph translating enterprise knowledge into action. The message: AMP provides the multiplayer rules; frontier models provide the muscle.

The most honest moment of the week came in a press briefing, where Dave Meyer, Atlassian's head of product, named the problem all of this is trying to solve. Models are getting dramatically smarter, he argued, and yet that intelligence is not showing up as proportional revenue gains at the companies building software. Teams ship more code while coordination gets harder: more quality problems, more review problems, more rework, more systems rebuilt because the plan was never fully captured.

That framing is the whole thesis of AMP: the bottleneck in AI-driven work is no longer the model, but the loss of context between people and their systems, and the difficulty of coordinating humans and machines toward the same plan. Atlassian is betting the company that owns the shared record of what a team decided, planned, and built has a structural advantage.

What to watch

AMP is a protocol announcement, and protocols are only as real as their adoption. The strongest signal Atlassian has is openness: by letting customers bring their own agents and models into the flow, AMP is positioned less as a product and more as connective tissue, the kind of layer that becomes hard to rip out once teams build on it. If third-party vendors build to it, Atlassian's suite becomes the room where human-agent work happens rather than just another tab.

The open questions are enterprise-flavored: will Non-Human Identities and Agent Sessions satisfy the compliance teams that block autonomous agents from touching production systems, and does the multi-model story hold up for third-party agents under Atlassian's governance rather than Rovo's? The answers will decide whether agents graduate from private chat windows to the shared spaces where the work gets done.