Atlassian launches AMP, a way for AI agents to work inside team tools like people

Moving AI out of the private chat

Atlassian, the maker of Jira and Confluence, announced AMP, short for Agentic Multiplayer Protocol, at its Team '26 Europe event in Amsterdam. As atlassian.com reported in a post signed by its chief, the pitch rests on a simple complaint: at many companies, AI starts as one person typing into a chat window on a laptop.

That can make one person faster, the post argues, but it does little for a team. The company's answer is to make AI "multiplayer", meaning agents (AI programs that carry out tasks) work in the same spaces as colleagues. Atlassian says rollout of AMP begins today.

Three layers of how AMP works

Atlassian describes AMP as three connected layers. The first is working where the work already happens. You can brief an agent with a short Loom video, mention one by name in a Confluence document, or argue over an approach in a Jira comment thread. Agents can also be plugged into Jira workflows to sort incoming requests or send work to the right place.

The second layer is identity and presence. Each agent has a named owner, its own profile and a live record of what it did. In a shared document or whiteboard, the agent's cursor appears beside those of the people editing. When a developer runs an agent session on a personal laptop, AMP is meant to surface that session on the team's Jira board so the work isn't hidden.

The third layer is context and governance. Agents draw on the Teamwork Graph, Atlassian's map of information from a company's tools. Permissions are limited by role: an HR agent acts with the access of whoever is prompting it, while a helpdesk agent uses a dedicated service account. Admins can see who triggered an agent, what data it touched and where human approval steps sit.

The supporting upgrades

Alongside AMP, Atlassian said it has shipped over 120 enterprise capabilities in the last year, with more at the event. Several are about feeding agents better information.

• MCP server: MCP, the Model Context Protocol, is a standard way for AI tools such as Cursor, Claude or VS Code to reach outside data. Atlassian says its updated server covers more and uses up to 25% fewer tokens (the chunks of text a model reads and writes, which drive cost) for the same Jira and Confluence work. That figure comes from internal benchmarking on Claude models.
• Code and data: The Teamwork Graph can now read source code down to individual functions, symbols and classes through a Code Search app. Company data warehouses can be queried too.
• Reach: The graph covers 7 context types and 250 billion connected objects, with 80+ connectors. New ones include Zoom, Gong, Microsoft Entra ID and Google Identity.
• Artifacts: a place to store what agents produce, under company controls, and feed it back into the graph.

Treating agents as accountable actors

The most practical announcement may be about control. Atlassian is adding agent accounts and an inventory of non-human identities. Each agent, app and service account, whether Atlassian's or a third party's, gets a full identity. Admins can manage these automated actors next to employees and cut off access when needed.

Atlassian also said it is rolling out local EU AI inference, which keeps language-model processing on models hosted in the EU. Another feature, tracking local agent sessions from developer tools and terminals inside Jira, is described only as coming "soon".

What to read into it

This is a company describing its own product at its own event, so the claims, including the 25% figure, are Atlassian's and have not been independently checked. The post doesn't give pricing, a full list of what is available now, or how customers have fared.

The underlying bet is clear, though. Atlassian expects most teams to have people and agents side by side, and it says the basics of teamwork stay the same: shared background, clear ownership and visibility into what is happening. Those are also reasonable questions to put to any AI tool your team uses, from any vendor.