Atlassian unveils Agentic Multiplayer Protocol for human-AI collaboration

AMP is intended to give every agent the data context needed to be useful and limitations to remain safe within the platform, including an identity assigned by administration that defines authority and scope. "This is a multiplayer game," Jamil Valliani, Atlassian's head of AI product, told SiliconANGLE in an interview. "Humans and agents are working together in a very dynamic space, and there are lots of players."

Agents are not meant to run without oversight. Valliani said Atlassian has an internal statement: "Headless software means brainless software." The company's role is not merely to provide software for autonomous agents to connect to and operate without people, according to the article. Autonomy is useful, but it is not useful without teamwork.

Atlassian announced Rovo Work, a new mode for Rovo Chat that handles complex, multi-step tasks that humans review and approve. It is designed for long-horizon tasks. "When you opt to send a query to Rovo Work, you're actually giving Rovo permission to go and actually unleash itself fully," Valliani said.

Rovo can also try to train itself when it encounters an unfamiliar task and lacks what it needs in its model training data. In one case, a product manager asked Work for an Instagram-ready reel. Rather than using whatever it had associated with its base model, Rovo researched the right instructions, learned what it needed for the output format, and obtained tooling for video synthesis. That approach differs from the standard AI model question of whether something is already available in training data to the question of how to learn to do it.

Valliani also confirmed that for power users, Rovo could generate custom skills as part of its work. That is a separately supported capability, rather than every learned behavior becoming a skill users can export. "We can't really imagine the creativity our customers are going to ask it to unleash," Valliani said. "We want customers to go and challenge it, and tell it what they really want."

Atlassian is also using Loom to bring visual instruction to everyday users, an approach developers have used with coding agents for a long time. Loom lets users record video of themselves and their computer screen and share it to provide instruction. It can also instruct agents: for example, drawing circles around parts of a user interface, showing buttons being clicked and information being entered into fields. The problem with AI agents is not always intelligence, according to the article. Sometimes humans have trouble explaining in words what they want, but they can visually point out what they need along with words such as "Make this larger, move this, turn it green." Two humans looking at the same screen while speaking have the same context, with a visual prompt plus words. "With Loom, it'll actually capture you saying all that, and the actual references on screen that you're pointing to when you say it, and then format it into the prompt," Valliani said.

For developers, Atlassian's broader platform changes expand the context and operating environment for AI agents. Rovo Code Search brings source code into the Teamwork Graph, a centralized data intelligence and context engine that maps relationships among people, code and documents. Data Context extends that view into structured information stored in platforms such as Databricks Inc., Snowflake Inc. and Google LLC's BigQuery. Together, Atlassian said, they give agents a wider picture of not only how software is built but also the business information surrounding why it is being built.

The company's expanded Model Context Protocol server also gives external coding and AI agents a common interface into its platform. Valliani said a substantial portion of those interactions now involve agents writing information back, not just retrieving it.