Atlassian has rebuilt its products so that roughly 95% of what a person can do in the user interface can also be done by a coding agent through a command line or an MCP connection, across more than 500 exposed tools.
The reason is that the customer is no longer reliably a person. A Jira ticket now gets read and written by Claude Code, Codex, Cursor or Copilot as often as by the engineer who filed it.
"They might be an agent coming in working with it."
Tamar Yehoshua ran product at Google Search and at Slack before becoming Atlassian's chief product and AI officer, and she joined the company at the moment the market decided agents would eat enterprise software.
The full interview is covered here so you can skip it. 54 minutes of audio, 15 minutes of reading.
Here are the 11 lessons that matter.
👤 Guest: Tamar Yehoshua, Chief Product & AI Officer at Atlassian, who previously led product at Google Search and at Slack
🎙️ Host: Michael Krigsman, industry analyst and publisher of CXOTalk, who takes audience questions live on LinkedIn and X during the show
📰 Published: 16 September 2026 on YouTube (CXOTalk) and the show's own feed
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 54 min | ✅ Time saved: 39 min
Key Takeaways
Atlassian rebuilt its products so an agent can do about 95% of what a person can do in the interface
More than 500 Atlassian tools are exposed to outside agents through MCP and a command line
Feeding a coding agent Atlassian's organization graph instead of raw document access raised answer quality 44% and cut token use 48%
The graph holds over 200 billion entities and is built from the customer's own tools
An agent's access is the access of the person who built it, and nothing more
A triage agent at Mercedes-Benz removed 85% of the manual work of sorting car-test defect tickets
The internal rule is to look for 10x process improvements, not 10% to 20% ones
One team used Figma's MCP to find gaps between design and code and fixed 14 bugs in an hour
Coding shrank from about 20% of the software lifecycle, and everything on either side of it got harder
Atlassian's answer is an AI planner before the code and an AI reviewer and site-reliability agent after it
Yehoshua does not believe agents kill software, and points at 370,000 customers and 85% of the Fortune 500
Her view is that vibe-coded internal replacements fail at scale because somebody has to support them
Information no longer has to travel up a hierarchy, and she has stopped holding monthly OKR meetings
An agent reads the documents, tickets and messages and marks each objective green or red
1. Agents as a Second User
Krigsman opened on what changes in product design when the user might be software. Yehoshua's answer was that the discipline does not change, but the entry point does.
Her framing of the shift is a return to fundamentals: "And now people are getting back to the same principles of product management of build a product that is great and that works." The early phase, she said, was the opposite — engineers looking at an interesting technology and asking what could be done with it.
"But the new twist is people may be using your product through a coding agent, through Claude Code or Cowork, through Codex, through MCP. So they're not always coming to your site, a human being coming and working with your interface. They might be an agent coming in working with it."
Atlassian's stated mission now includes non-human members: "Our mission is to unleash the potential of every team. And now that team includes agents, humans and agents."
The engineering target was parity between the interface and the programmatic path. For Confluence, she said, "we've gotten to the point where 95% of everything you can do in the UI, you can do through MCP or a CLI so that you can access that." The same goal applies to reading and writing Jira issues from a coding agent.
The surface exposed to outside agents is large: "So like our MCP and CLIs have over 500 tools from Atlassian that we expose to the agents."
The constraint she flagged is that more context is not better. Too much confuses the model and costs money, so the judgment is which tools to expose as much as which data.
2. The Teamwork Graph
Asked what "the right context" actually means, Yehoshua described the product Atlassian has built to supply it.
"We have something called the teamwork graph. So the teamwork graph is the context graph that we build for each customer for their organization. And just by using our tools, it builds your graph." It predates the current AI cycle; investment in it has risen sharply since.
It is a map of the organization — people, knowledge, structured data, code and communication — and it reaches outside Atlassian. She named Google Docs, SharePoint, email, Slack and Teams messages, with Databricks, Snowflake and Google for structured data and Workday for people data.
Scale: "Our teamwork graph has over 200 billion entities in it, and there are links between it." The slice any individual sees is bounded by their own permissions.
The advantage over plain document retrieval is precomputation. Asking which projects a team is working on resolves through hierarchy and authorship already stored in the graph rather than a broad sweep of everything a user has ever touched.
The measured result is the number that carries the section. In A/B tests running Claude Code or Codex with MCP alone against the same agent with the teamwork graph, "the quality increased by 44%. And very importantly, the token usage went down by 48%."
Atlassian exposes the graph rather than guarding it. It powers Rovo, Agent Studio and Atlassian's automations, and is available to third-party coding agents through MCP and the command line. "So we believe in a very open platform."
3. Permissions, Not Prompts
An audience question from Arsalan Khan asked whether agents should be onboarded with deliberate access, and who decides. Yehoshua's answer is that the question is already settled by identity.
"The way that the context works in our products is it's all driven by whatever permissions the human has." An agent is an extension of the person who built it and cannot see a private conversation that person cannot see.
Third-party agents calling in from outside Atlassian also call using the builder's identity, with what she described as a lot of safeguards built around it.
Rovo also accumulates memory of what a user has done, which she said makes it more useful over time. Atlassian has onboarding agents for people and has not yet built one for an agent; in her experience, pointing a new agent at the right thing for the job has been enough.
On a question from Smail B. about data provenance and human-in-the-loop triggers, she named a product: "We have a product called Guard in our enterprise offering that does a whole data sensitivity layer." Rules can exclude classes of data outright or force checks.
Where the human stops the process is a configuration choice, not a product default. Atlassian's products run millions of automations a month, workflows that fire on events like a bug of a certain type arriving; a workflow step can call an agent and then require a person to sign off.
The control she puts most weight on is visibility. An administrator can see every agent built, decide who is allowed to build, and watch token spend against value. Some customers restrict building to a trained group in IT or engineering; others open it to everyone and monitor.
Her summary of the design goal is an open platform with an enterprise off switch — flexible by default, with the controls to "turn the spigot on or off."
4. Bug Triage Down 85%
Asked for something concrete, Yehoshua used a customer example rather than a product feature.
Mercedes-Benz has used Atlassian for over a decade, and files a Jira ticket every time a test engineer finds a defect in a new car. The volume produces duplicates and incomplete tickets.
"So What they did is they built a Rovo agent that takes in all of those tickets and auto-triages them, removes the duplicates, knows who to assign them to, and just gives all the details that you need. And so that triaging agent reduced the manual work for engineers to triage the bugs by 85%."
Her point in telling it is that customers build things the vendor never imagined, and templates exist to get them started rather than to constrain them. "And we want people to build stuff that we've never imagined."
5. Aim for 10x, Not 20%
A question from Ricardo Anklan asked how leaders balance urgency against picking the right tasks. Yehoshua's answer was about goal-setting and measurement.
She has seen the unstructured version — an instruction to go and use AI — and said it has real value as learning, provided the outcome is measured. Without measurement, speed hides the absence of value.
Atlassian runs structured internal experiments with goals per department, in engineering, finance and sales, built on Rovo. The company describes itself as its own day-zero customer and publishes a report on it.
The bar is deliberately high: "And we ask people to think about what are the processes where you think you can get a 10x improvement. So this isn't where you want just an incremental 10 to 20% improvement."
Coding agents alone, she said, only get a team so far. The instruction to one team working on Confluence slides was to reimagine the whole workflow.
The example she gave from that team is narrow and concrete: they used Figma's MCP to find the differences between the design files and the shipped code, then used coding agents to close them. "So they fixed 14 bugs in an hour, which it would have taken days."
Her closing advice on the question is that a quarter or a month later the goals tell you which teams used the tools well — and that the teams that missed are also information.
6. Efficiency Before Novelty
Krigsman asked whether agents are being used mostly to do old things faster or new things at all. Yehoshua said the sequence is always the same.
Her analogy is the first decade of mobile: the earliest apps were websites on a phone, and only later did anyone use location and maps the way Uber did. Enterprise AI is at the equivalent stage.
The example of the second kind is Confluence Remix with Rovo, which lets a user highlight text on a page and turn it into a different format. She said that is something customers could not previously do at all, and that the team built it in weeks rather than months while inventing new evaluation methods to do it.
Her read is that the models are further ahead than the organizations using them. Companies started with chat and search and are only now redesigning workflows.
On the standard objection that large companies are risk-averse, she said this cycle looks different to her. Every customer she meets asks what Atlassian is doing internally, and the motive is partly fear of competitors moving faster.
What unlocked adoption on the buyer side was enterprise assurance rather than persuasion. She named zero data retention as the kind of commitment that turned reluctance into what she called a huge uptick in usage.
7. Tokens vs SaaS Licenses
Michael Beelar asked from LinkedIn how Atlassian looks at its budget across software licenses and tokens. Yehoshua rejected the trade-off.
"Value-driven." Almost all of Atlassian's internal token cost sits in engineering; usage elsewhere is small by comparison.
The metrics are output metrics: pull requests deployed per engineer, features developed, and whether objectives are met. She said both PRs and features are up sharply.
Atlassian acquired DX last fall, which measures AI adoption inside engineering teams and adds survey data on whether engineers feel more productive.
She does not see licenses and tokens as substitutes: "So that I don't feel like it's a trade-off between the two, because if you have a need for those products, those needs don't go away." Atlassian still buys Databricks and Salesforce, and now uses their AI features too.
Her team also runs corporate development and Atlassian Ventures, which she uses to keep track of new AI products both to build with and to buy.
Cost control at the product level runs through routing rather than negotiation. Atlassian built an AI gateway that picks a model per job on quality, latency and cost: "So we do not use one model in our products." Most work does not need the frontier model, and prices for the previous generation keep falling.
8. The SaaSpocalypse Call
Krigsman put the bear case directly: agents destroy software. Yehoshua said she had never believed it, and joined during the panic.
"I joined Atlassian right around the time of the SaaSpocalypse, and people are like, wait, what do you mean you're going to a SaaS company? And I just never believed the narrative, partially for what I said."
Her evidence is installed base and criticality: "We have 370,000 customers. We are in 85% of the Fortune 500. People run their tier 0 workflows on Atlassian."
The conditional matters, and she stated it: "Now, if we did nothing, And didn't introduce AI into our products, I would be worried." What reassured her before joining was the infrastructure already built.
Asked whether companies will simply build their own versions, she drew a line at company size: "I think that works for small companies. I don't think it works for large companies." Past a few hundred people, whoever wrote the internal tool becomes a software vendor with support obligations, single sign-on, data protection and a queue of internal customers.
"We spend a lot of our engineering resources on enterprise readiness."
Inside Atlassian, the building that does happen is agents rather than replacements: agents that prepare sales calls and chase approvals, agents that help finance close the books, and the largest concentration in engineering. Product managers, she said, no longer write their weekly updates.
9. Left and Right of Code
Swami Vaidyanathan asked from LinkedIn what happens when agents chain into other agents and no human reviews the intermediate steps. Yehoshua said the industry has not solved it and named the principles she works from.
"I don't think anybody has figured out the answer to this. So I'm not going to claim to have figured it out."
The accountability rule is unambiguous: "whoever built the agent, the agent that starts is still accountable. There always has to be a human accountable."
Her expectation is that the checking gets automated too: "And I also think that what's going to happen is agents are going to be putting in guardrails for other agents." Atlassian already uses a model to judge another model's output when testing features.
Krigsman brought a data point from the previous week's guest, the head of engineering at Snap: half of that company's engineering investment goes into the agent-assisted engineering and half into guardrails and accountability. Yehoshua agreed flatly: "100%."
Her structural version of the same point starts from a number: "So, if you think of pre-AI, if you think of the software development lifecycle, About approximately 20% of your time was spent coding." Coding has shrunk; the work before and after it has grown and become more complicated.
The mechanism is volume. If the rate of incidents per pull request holds steady and the number of pull requests rises sharply, incidents rise with it.
Atlassian's response is a tool at each stage. An AI Planner turns a prompt into a spec, a technical document and work items, using the context graph to pull in prior code, design documents and the people involved. An AI reviewer checks generated code against the organization's own past incidents, which are documented in Jira. An AI site-reliability agent watches for the issues that have previously appeared in production.
10. What the UI Becomes
Eldad Postan-Koren asked which parts of today's software architecture go obsolete first if agents become users.
Her answer is that interfaces get simpler rather than disappear. The complicated Jira forms an administrator configured are the first casualty, for people as well as for agents.
The new design problem runs the other way: humans need to see what the agents are doing. "So how do you visualize when a project used to take 10 people, 50 people, and now you have agents and you may have hundreds, maybe thousands of agents running on something."
She flagged the conclusion as contrarian: "We're investing a lot in design because we, this might be a contrarian view, but we actually feel that the design of products is gonna become so important of how do you understand and observe what's going on." Design becomes a differentiator rather than a cost.
Atlassian built teamworkgraph.com as a demonstration of the principle — a site where a customer can log in and see their own graph, who they are connected to and how, because a system people are asked to rely on has to be visible.
11. Leading in the AI Era
Tim Crawford asked, through the chat, where the efficiency work leads. Yehoshua's answer was about how organizations move information.
"So if you think pre-AI, information flowed through hierarchies." With Rovo, anyone in the company can ask the status of an objective or what caused an outage without going up a level.
She has stopped holding a recurring meeting because of it: "I, for example, no longer have monthly OKR meetings. I have an agent that runs, that gives me the updated OKRs, goes through, looks at all the Confluence docs, the Jira tickets, the Slack messages, and then says these are green or these are red." She then calls a meeting only on the red ones.
Her claim is that this is a leadership problem, not a tooling one: "And I think all leaders need to change how they're thinking about how they run organizations and where their points of leverage are." Without that, she said, the gains stay at the level of micro improvements.
The example she gave of leadership depth is her own chief executive. Mike Cannon-Brookes became absorbed in OpenClaw when it launched, built agents to run his personal life, and told the executive team to implement it themselves so they would understand what people were talking about.
The internal training program is AI Builders Week, once a quarter, in which product and design stop other work for a week to learn a subject — prototyping, evaluations, agent building or coding — and build and demonstrate a project.
The measured result: "70% of what's built during Builders Week actually gets used later."
Her reason for running it is time rather than instruction. The complaint she hears most is that people are too overworked to learn the tools, and she said the space has to be given rather than assumed.
Asked for advice to chief information officers, she gave a pairing: "So be ambitious, but be educated." Do not set a goal the tools cannot reach, but do not assume the tools can do less than they can.
Bonus Insights
On starting small rather than everywhere: "So one issue I think people do is they go too large with AI. I'm going to transform everything, and those usually fail." Her sequence is to transform one business process, learn the capability, then set the ambitious target.
On the Forge marketplace, she called partner access to Atlassian's AI a critical part of strategy. Apps built on Forge can call large language models through Atlassian's AI gateway and use its agents.
Rovo agents can work across projects, she confirmed to an audience question, subject to the same access rule: whatever the person who built the agent can see.
Krigsman used part of the hour to trail forthcoming guests — the chief AI officer of Occidental Petroleum, the vice chair of Deloitte, and a pending booking with the chief AI officer of Bank of America.
Yehoshua's bottom line is that agents do not replace enterprise software so much as change who logs in: the product still has to be worth using, but it now has to expose everything through a programmatic path, carry the organization's context, and show a human what several hundred agents did on their behalf.
Products, Companies & Tools Mentioned
Atlassian (Yehoshua's employer, with 370,000 customers and a presence in 85% of the Fortune 500, rebuilding Jira, Confluence and the rest for agent access)
Jira and Confluence (The products used throughout as examples, both now readable and writable by outside coding agents)
Rovo (Atlassian's AI product and agent-building studio, and the basis of the Mercedes-Benz triage agent)
Mercedes-Benz (A customer of more than a decade, whose Rovo triage agent cut manual defect-ticket work by 85%)
Figma (Its MCP connection was used by an Atlassian team to find gaps between design and code, fixing 14 bugs in an hour)
Databricks, Snowflake and Workday (Sources Atlassian reads into the teamwork graph for structured data and people data)
Slack, Microsoft Teams, Google Docs and SharePoint (The communication and document systems the graph ingests alongside Atlassian's own)
Claude Code, Codex, Cursor and GitHub Copilot (The outside coding agents Atlassian treats as users of its products)
DX (Acquired by Atlassian last fall to measure AI adoption and developer productivity)
Snap (Cited by Krigsman from a prior episode: half its engineering investment goes into guardrails and accountability)
Salesforce (Named alongside Databricks as software Atlassian still buys, as her evidence that tokens are not replacing licenses)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: