Mike Cannon-Brookes on AI Agents, Enterprise Context and the Future of Interfaces

Atlassian co-founder and CEO Mike Cannon-Brookes joins Harry Stebbings to discuss how enterprise software changes when agents can retrieve information and act. His central argument is that agents will transform many tasks without eliminating the need for thoughtful interfaces, company context, and human judgment.

Stebbings asks whether software providers risk becoming mere databases as agents perform work for users. Cannon-Brookes rejects the premise as too broad: the answer depends on the kind of work. A highly functional task is different from one that relies on taste, interaction, and collaboration.

He compares buying a replacement charging cable with shopping for something where the experience and research matter. An agent may remove the tedious steps in the first case, while a person still wants to participate in the second. Enterprise work likewise contains both routine steps and creative decisions.

Agents can move collaboration and business workflows forward, giving people more time for higher-value creative work, Cannon-Brookes says. Jobs will change, but he sees that as an opportunity for Atlassian. The actual design of enterprise experiences around agents remains far from settled.

His shorthand is that 'headless is brainless': giving up on improving the interface and customer experience is a poor bet. He does not mean rejecting machine access. Atlassian invests in MCP servers and command-line interfaces so agents can read from and contribute data to its platform and Teamwork Graph.

That agent access also reveals the value of a context graph, he argues. But if a company's own interface and design stop evolving to solve meaningful customer problems, it risks losing its place in the experience. Human-facing design and machine-facing access can both matter.

Cannon-Brookes believes design and experience become more important as creating technology gets cheaper and more abundant. An application that looks excellent today can feel dated in five years as expectations rise. Continuously improving the experience takes research, development, and craft; it is not a one-time design exercise.

Stebbings asks whether horizontal agents will instead span multiple vertical providers. Cannon-Brookes thinks horizontal agents, deeply specialized agents, and interfaces can coexist. Agents in products such as Claude, Codex, Cursor, Teams, and Slack can interact with Atlassian without becoming the best interface for every Atlassian task.

He describes a Confluence user selecting text and asking for a number to be checked against information in Databricks or Snowflake, perhaps receiving a small chart in place. That quick interaction is useful without turning Confluence into a full business-intelligence tool. The right experience depends on whether someone needs an answer or is immersed in a specialist workflow.

The same principle applies to Figma: an agent in Atlassian can reach Figma functionality, and people can access Atlassian from Figma. Cannon-Brookes sees value in agent-to-agent interactions across teams, while each interface retains its own strengths. He is skeptical of a single magical button or universal agent that does everything.

Among Atlassian's important strategic choices, he names investing in its enterprise context graph and continuing to evolve design for the agent era. He compares today's foundational design challenges with the shift to mobile: teams must find new ways for people and agents to participate in collaborative work.

The questions include how agents appear, how humans interact with them, and what it means for an agent to act on someone's behalf. Provenance and accountability matter, but not everyone needs to see every agent trace. A good experience must support speed and understanding while allowing people to inspect details when necessary.

One unresolved problem is the blank text box. A single line of text may expose powerful capabilities, but most people do not know what to ask and often use the tool as a search engine. Designers need to make relevant capabilities discoverable and available at the right moment.

A second problem is attribution. If Codex creates a pull request for Harry, should it be credited to Harry, the agent, or the agent acting for Harry? The answer can depend on how involved the person was. Some readers may want human comments, others agent comments, and both can have legitimate uses.

Cannon-Brookes distinguishes asking AI to write a 3,000-word essay from asking it to improve an essay a person has already written. A simple watermark may describe both as AI-touched even though the contributions differ. He argues that provenance is a nuanced design problem rather than a rule that all AI assistance is bad.

He also questions the limits of the software-factory analogy. Software is creative, even if machines increasingly do parts of the work. People still need to initiate, steer, co-create, and handle exceptions. Interfaces should help them know when intervention is needed without burying them in a constant stream of agent activity.

Stebbings predicts that human-created content may eventually be the material that gets labelled. They recall how quickly once-unusual online behavior became normal. Cannon-Brookes remembers Atlassian's early customers downloading a form, faxing their credit-card numbers, and having staff enter payments into a physical machine.

Cannon-Brookes thinks people may become less concerned with whether content received some AI assistance, just as moviegoers usually judge a film by its story rather than whether it used color grading or visual effects. He still distinguishes a spectrum: entirely synthetic low-quality material may make human craft and judgment stand out.

Stebbings asks about enterprise reluctance to use Chinese open models. Cannon-Brookes says some customers have pushed back on models described as Chinese, while he does not see equivalent resistance to open models in general. He says Atlassian uses Nvidia and Mistral open models and, at the time of the interview, does not use Chinese models in production.

He says Atlassian's job is to select and combine models for customer outcomes, price, and performance rather than to favor one model. He also raises the complexity of model provenance and distillation, suggesting that the boundary between models from different origins may be harder to draw than customers assume. These are his observations and predictions, not an audit of individual models.

Cannon-Brookes uses Linux as an analogy for a globally contributed open-source project and separates strategic geopolitical control from an enterprise's choice of a particular open technology. He expects resistance to some open-weight models could change over time, while acknowledging that customers may choose not to use a given technology.

Stebbings raises concerns about the relationship between model developers and government. Cannon-Brookes declines to make a broader social judgment and returns to the technology choice: in his view, a cheaper open alternative that performs well may become difficult for enterprises to rule out indefinitely. He expects some organizations to adapt open models for their own uses.

Asked whether Atlassian should build its own model, Cannon-Brookes says the benefits of specialized models remain unsettled. They can make sense when an application is unusually specific or there is substantial proprietary data, such as in some robotics or world-model work. He is less convinced the benefit will persist for every application.

A specialized model can have a price-performance advantage at launch, but general models may improve quickly enough to erase it. He draws a distinction between training a model from raw data and adapting an existing open-weight model. Atlassian fine-tunes models for specific purposes, he says, but he does not describe that as training its own model from scratch.

What would change his mind? Sustained evidence that specialization stays ahead of the latest general foundation models on the relevant price-performance measures. The answer depends on the application, latency, cost, and timing. Improvements in an agent harness can make a model more useful, but he distinguishes harness design from progress in the underlying model.

A durable specialized advantage, he argues, requires both an ongoing stream of useful specialized data and a problem general models do not simply absorb. He says some startups that built models a few years earlier have seen general models improve faster than their ability to preserve a specialized lead.

Stebbings contrasts well-funded private companies that spend freely on frontier models with the pressures facing a public company. Cannon-Brookes says cost discipline matters over time. He considers it valuable to learn how to apply a cheaper or older model effectively, rather than relying on the quickest result from the most expensive model.

He says large enterprise customers increasingly want to understand token spending: which models and features their teams use, what value they receive, and where spending can be optimized. Whether they purchase a model directly or pay for software with AI costs embedded in its price, that expenditure ultimately matters.

Model routing and blending are part of the answer: choose the right model and harness for a task while balancing quality, cost, and speed. Cheaper decision models, he says, can make it viable to add probabilistic capabilities to features that would have been too costly with earlier models.

For a company at Atlassian's scale, Cannon-Brookes expects routing to be an important internal capability, potentially layered with external tools. Smaller businesses or firms without deep technology expertise may benefit from existing gateways. He cautions that automatic routing is difficult without enough direction or intent, and notes that agent harnesses can also route among models.

Atlassian's own AI gateway team handles routing for products such as Rovo, he says, seeking the best balance of answer quality, price, and speed. Because the platform operates at scale, the company also wants to keep reducing the cost of heavily used features.

His advice to product teams is not to optimize cost too early. First build a feature customers find valuable and consider whether its economics could improve. If five people use it, the cost may not matter much; if five million do, the team should work seriously on optimization as adoption grows.

Asked what will become much cheaper within five years, Cannon-Brookes points to model inference. He expects investments in chips, data-center design, model distillation, and other techniques to improve price and performance over time. It is a forecast based on the amount of talent and capital entering the field, not a guarantee.

Stebbings argues that lower prices may also cause people to use far more inference. Cannon-Brookes agrees usage could expand as costs fall, while maintaining that there will still be economic limits. His broader concern is that technology production is already moving faster than customers can realistically absorb changes into their businesses.

For people using models in coding, architecture, business processes, or art, he sees cheaper and more capable technology as positive. He credits Culture Amp's founder with a phrase he likes: the economy is moving from information asymmetry toward imagination asymmetry. Knowing something others do not may matter less than imagining a distinctive application of available technology.

Stebbings asks where Cannon-Brookes himself has not shown enough imagination. His answer is time: despite working long days, he wants more uninterrupted space for deep thinking. A fragmented stream of messages, short-form content, and agent activity can make it easy to remain busy without noticing slower, year-to-year changes.

Over the previous few months, he says, he has tried to protect even a few hours a week for that reflection. As more agents run tasks and compete for attention, creating space to think becomes a leadership skill. The two joke that Stebbings's own automated tasks include shopping for his girlfriend.

Stebbings turns to the market for data and asks about the cost of acquiring information for a context graph. Cannon-Brookes distinguishes large generic data sets sold to model builders from the knowledge accumulated inside a particular company. His focus is the latter: decisions, projects, documents, people, and code in the context of work.

Important facts do not always live neatly in databases; they may be buried in presentation slides or documents. Atlassian's aim, he says, is to connect organizational knowledge and history with relationships between people and technical services, then use that context to help customers and their agents find answers and learn from decisions.

He compares the Teamwork Graph's role to a search engine indexing the web before a query arrives. The system should prepare and infer useful connections ahead of time so it can answer efficiently, but each company's index has to reflect its own distinct information. He does not assume that companies will sell this internal context as a generic data product.

The graph must also remain fresh. People join and leave, projects change, and working relationships shift. Cannon-Brookes distinguishes current facts from historical learning: a system can retain useful lessons from an organization's past while recognizing that many old details no longer describe what people do today.

His own work is an example. The Java code he wrote in 2003 is not very useful to Atlassian today, but knowing he once programmed may help someone explain a technical concept to him. His present work is management, storytelling, and some product and design work; a useful graph should represent that change.

He contrasts company context with specialized training data for self-driving cars. A car needs to recognize roads, signs, cyclists, and other features of its environment; it does not need every fact in a general encyclopedia. Different applications create different requirements for what information is valuable.

Asked which company he would acquire for its data, Cannon-Brookes says acquiring data is not the main point: the context belongs to customers. Atlassian aims to connect their information, offer lexical and semantic search, and provide access through interfaces such as MCP and the command line. Customers can add data specific to their industries alongside common materials like meetings and documents.

Stebbings asks how Cannon-Brookes views OpenAI and Anthropic moving into applications such as documents. He says both are important Atlassian partners and that the eventual boundary between model-provider services and applications is contested and uncertain. Companies can be partners and potential competitors at the same time.

He points to Confluence as an example of value beyond the ability to enter text. He says its last disclosed annual recurring revenue was above $1.5 billion and that it was growing more than 25% annually at the time. Many tools can store text; design, purpose, user experience, and integration with other workflows help make a product useful.

Stebbings asks whether Linear is perceived as showing more design craft than Jira. Cannon-Brookes calls that comparison unfair and stresses sustained improvement: standards change, so software teams must repeatedly raise their level. He points to investment in Jira and the Teamwork collection's speed, performance, and experience, and says Atlassian focuses on the long race.

The conversation turns to young founders selling early and the energy required to stay for decades. Cannon-Brookes says enjoyment and managing one's energy matter. He enjoys working with smart people on difficult problems, and although his co-founder Scott is no longer in the same daily role, they still meet regularly and discuss Atlassian, software, and AI.

He wants his children to find the activities that give them energy. Founding a company can be especially stressful when fundraising, employee livelihoods, revenue targets, and personal reputation all feel intertwined. He credits founders for taking on the challenge whether their companies ultimately succeed or fail.

Money and status alone, he argues, are rarely enough to sustain someone through five or ten years of building. Enjoyment of creating a product or service and working with people can supply a more durable source of motivation. He has seen waves of people arrive for a fashionable technology and leave as the fashion changes.

Stebbings worries that venture capital has made it too easy for some founders to raise large rounds, play the founder role, and exit quickly. Cannon-Brookes describes capital allocation as a market outcome, but returns to his own source of energy: sustained work with people he respects on difficult problems.

In a quick-fire round, Cannon-Brookes says he has revised his expectation of future inference prices downward. Asked what people overlook about wealth, he calls it both wonderful and a burden: he sees it as a temporary stewardship that carries responsibility to do things beneficial to society, not only personally profitable things.

Asked how to raise children well amid wealth, he offers humility, curiosity, and gratitude as qualities parents can model, while admitting the challenge and his own uncertainty. Stebbings adds that children watch how adults treat others. Both connect the lesson to behavior rather than simply telling children what values to hold.

The advice Cannon-Brookes returns to is to 'wage peace on yourself.' Technology workers can set extremely high standards and create stress for themselves, he says. When you have genuinely tried and done the right things, he recommends not being unnecessarily harsh with yourself over the outcome.

For his own future growth, he returns to the need for slow thinking. Notifications and social media already fragment attention; faster technology and more agents could intensify that pressure. He wants leaders to create moments away from the screen and make time to think deeply, even as the world seems to move faster.

Stebbings closes by thanking Cannon-Brookes for helping others without expecting an immediate benefit. Cannon-Brookes offers to continue helping. The exchange reinforces the episode's final theme: the value of relationships, care, and attention alongside the rush to build new technology.