Atlassian AMP Agentic Multiplayer Protocol graphic showing agent presence, identity, attribution and real-time editing — Atlassian announced AMP, the Agentic Multiplayer Protocol, at Team '26 Europe on October 7, 2026. Image: Atlassian.

Ask most engineering leads a simple question: how much of your codebase is AI-generated code? Very few can give you a real number. Atlassian thinks that blind spot is now the main thing slowing AI down inside big companies.

On October 7, 2026, at its Team '26 Europe event in Amsterdam, Atlassian announced AMP, the Agentic Multiplayer Protocol. It gives every AI agent an owner, an identity, limited permissions and an audit trail. It also promises version history that shows what a human wrote and what an agent did.

I'm Nitin from Debugged Pro. Our developers use coding agents every week on client projects for teams in India and the USA. So the question "who wrote this?" is not theory for us. Here is what Atlassian launched, what experts doubt, and what I think agencies should do right now.

What Atlassian Announced on AI-Generated Code

Atlassian calls AMP its "foundation for how humans and agents collaborate across the Atlassian platform." In its official Team '26 Europe announcement, the company says AMP defines how agents take part in team work with an identity, scoped authority, shared context and results you can review.

In plain words, an agent stops being a hidden helper in someone's terminal. It becomes a visible teammate with a name, a boss and a set of rules.

Atlassian groups AMP around three ideas:

Work in the flow. Agents can be pulled in through @mentions in Confluence, Jira comment threads or Loom video briefs. The Atlassian MCP Server brings in context from outside tools like Figma or IDEs.
Identity and shared presence. Every agent has a clear owner and its own profile. Agents show up in presence bars and cursors next to human teammates. Local agent sessions on a developer's laptop link straight into Jira boards.
Context and governance. Agents draw on the Teamwork Graph, which Atlassian says now connects more than 250 billion objects and relationships. Agents can "Run as User" or use a dedicated service account, with a full audit trail of every action.

CEO Mike Cannon-Brookes put the pitch simply: "The best work has always been done in teams, and that's true of AI too."

The Parts That Touch Your Codebase

For developers, a few pieces matter more than the rest:

Rovo Code Search. The Teamwork Graph now reads source code down to functions, symbols and classes. Engineers can search across Bitbucket and GitHub in plain language without cloning a repo.
Attribution in version history. According to InfoWorld, Atlassian says version history will show exactly what a human wrote versus what an agent did, across Claude, Codex, Figma and Rovo.
Agent sessions in Jira. Local and cloud agent runs link back to work items. So the "why" behind a change stays attached to the ticket.
Interactive PR reviews. Agents can record a Loom walkthrough of their own code changes in Bitbucket. They explain what they did and the key tradeoffs for the human reviewer.
Rovo Work. A new mode in Rovo Chat for long, multi-step tasks. You set a goal, it proposes a plan, you approve it, and it can run for hours in a sandbox your admins control.

Atlassian also shared some usage numbers. Its MCP server has nearly 2 million monthly active users and handles over 15 million tool calls a day. The company says that is a 15x jump in six months. People and agents now work together on the platform more than 10 million times a month.

What Is Live and What Is Still "Soon"

This detail matters a lot. InfoWorld reports that Rovo Work, the code attribution features and EU-only AI processing are available now. The deeper non-human identity controls are not.

Atlassian says controls over what AI can see, plus limits on what agent accounts can access, will roll out "soon." Its blog adds that admins will get one inventory of every agent, app and service account, with the option to revoke access with a single switch.

Atlassian Administration screen listing non-human identities such as agents and apps with status, actor type and mode — Atlassian's planned non-human identities view lists every agent and app with its status and mode. Atlassian says these controls are coming soon. Image: Atlassian.

Why Tracking AI-Generated Code Is So Hard

If you think Git already solves this, it does not. The best part of InfoWorld's report by Evan Schuman is how clearly the experts explain the gap.

Adam Resnick, a research manager at IDC, pointed out that Git records an author and a committer. "So when a developer runs an agent locally, the record names the developer," he said. Attribution usually lives in commit message text, which developers can turn off, edit or lose.

He listed more problems. Each vendor tags commits in its own way. Developers edit agent output, so human and AI work gets mixed. And many teams run several agents side by side.

Mike Wilkes, enterprise CISO at Aikido Security, described the same issue from the security side. "Git can tell you who committed a change, but that is increasingly different from who, or what, actually wrote the code," he said.

Jamil Valliani, head of AI products at Atlassian, told InfoWorld that poor visibility into code origin is a massive problem. "For most enterprises, this is what is holding them back from further adopting AI," he said.

The Hard Middle

Frank Dickson, principal analyst at Dickson Research, gave the most useful examples. An agent using a person's login gets recorded as that person. A developer who accepts an AI suggestion signs it as their own work. A human who rewrites three lines of agent code creates a change that belongs to both.

"Attribution is easy at the edges; it is hard in the middle, where most of the work now happens," he said.

That matches what we see. On a typical feature, a developer prompts an agent, rejects half of it, fixes the rest and adds tests by hand. Labeling that commit "AI" or "human" tells you very little.

What the Experts Doubt

Analysts liked the direction but were careful about the details. Here are the three main concerns from the InfoWorld piece.

Labels alone are thin. Dickson said the labeling is a weak differentiator on its own. The real value sits in the identity control layer, and that is the part marked "soon."
Setup is not approval. Atlassian says it traces each change to the agent and "the person who set it up." Dickson noted that the person who set up an agent may not be the one who approved its work. "The person who approves an agent's work owns it," he said.
Who typed it matters less than who judged it. Resnick said companies want to see where humans made the big calls, like which data store to use or how to handle login. The useful record shows real human review, "rather than simply recording that someone clicked approve."

Justin Greis, CEO of consulting firm Acceligence, added a practical view. You may not need line-by-line provenance to judge a tool. You can compare cycle time, defect rates, rework and security findings before and after a team adopts it.

What AI-Generated Code Tracking Means for Agencies in India and the USA

For a software agency, this is not just an enterprise IT story. It changes the questions clients ask us before they sign.

US clients already ask us how we use AI tools on their code. Indian enterprises and Global Capability Centres are starting to ask the same thing. Soon "we use AI responsibly" will not be enough. They will want proof.

Here is where I think AMP and tools like it will matter most:

Client trust. A clear record of which changes came from agents makes code handover easier. It also makes audits less painful.
Review effort. Resnick suggested risk-based review instead of one blanket rule. A small, well-scoped agent change may need a lighter check than a human edit to a payments system.
Real ROI numbers. Wilkes said good attribution lets leaders measure agent work instead of just counting licenses and tokens. That is exactly what clients want to see on a monthly report.
Atlassian shops get it first. If your client runs Jira, Confluence and Bitbucket, these features land inside tools they already pay for. We covered the wider Atlassian AI push in our post on Jira AI and the Atlassian OpenAI deal.

Diagram of the Atlassian MCP server connecting apps like Jira, Confluence and Google Drive to an AI chat that finds work items — The Atlassian MCP server links outside AI tools and coding agents to Jira, Confluence, Loom and Bitbucket. Image: Atlassian.

A Simple Plan to Track AI-Generated Code Today

You do not need to wait for Atlassian's "soon" features. Most of this works with any Git host and any coding agent. Here is the plan we follow on client work.

Give agents their own accounts. Do not let agents act under a developer's personal login where you can avoid it. A separate service account makes every agent action easy to spot and easy to revoke.
Name a human owner for every agent. Write down who set it up and who approves its output. As Dickson noted, those may be two different people.
Add one line to your PR template. Ask "Which parts were agent-written, and what did you change?" It takes ten seconds and gives reviewers context.
Use commit trailers consistently. Pick one convention, such as a co-author line for the agent, and enforce it in your repo rules. Mixed conventions are the problem Resnick described.
Review by risk, not by source. Set stricter review for auth, payments and data handling, whoever wrote the code. Use lighter review for small, well-tested changes.
Track outcomes monthly. Compare cycle time, review time, rework and bugs for agent-heavy and human-heavy work. That turns opinion into data.

If you want a wider framework for AI risk, the NIST AI Risk Management Framework is free and widely used. It gives you plain categories for governance, mapping, measuring and managing AI risk.

For the review step itself, our guide to AI code review benchmarks and tools covers how to test whether an AI reviewer actually catches real bugs.

A Quick Look: Before and After AMP

Question	Plain Git today	Atlassian with AMP
Who wrote this change?	Usually the developer's name	Version history splits human and agent work
Which agents are active?	Often unknown	Owner and profile per agent (full inventory "soon")
What can an agent access?	Whatever the user can	Run as User or a service account, with audit trail
Why was it changed?	Commit message, if written	Agent sessions linked to Jira work items

Our Take

I like the direction. Treating agents like teammates with names, owners and limits is far healthier than letting them run unseen in a terminal. The usage numbers also show that teams are already working this way, whether or not they track it.

But I agree with the analysts. A label that says "agent wrote this" is the easy part. The value is in knowing who approved it, what the agent could touch, and whether the review was real. Much of that is still on the "soon" list.

So my advice is simple. Use what is live if your clients are on Atlassian. Build your own lightweight tracking either way. If you need help setting up agent workflows or review rules for a project, our software development consulting team can help you plan it.

Key Takeaways

Atlassian launched AMP on October 7, 2026, to give AI agents identity, owners, scoped permissions and audit trails.
Version history will split human and agent work across Claude, Codex, Figma and Rovo, per InfoWorld.
The strongest controls are not live yet. Non-human identity inventory and AI access limits are marked "soon."
Git alone cannot track AI-generated code well, because agents often commit under a developer's name.
Agencies should act now with agent service accounts, named owners, PR template questions and risk-based review.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with development teams and clients in India and the USA to plan software and AI projects, set delivery processes and keep code quality high. He writes about practical AI adoption for agencies and growing businesses.