Atlassian’s Cannon-Brookes: Agents Won’t Replace Interfaces
Summary
- A central product bet for Mike Cannon-Brookes is that agents will expand, not erase, the value of interface design. Functional “just get it done” work can disappear into agents, but collaborative work still needs taste, context and human interaction; horizontal agents, vertical agents and first-party interfaces can coexist. His blunt rule: “Headless is brainless,” and he is “a big bear on the magic button thesis of AI.”
- Atlassian’s strategic asset is the Teamwork Graph: a fresh, company-specific index of people, decisions, documents, code and workflows that agents can query. Its job is to infer and organize context before an agent asks, much as Google indexes the web before a search. Cannon-Brookes stresses, “It’s the customer’s context. It’s not our context,” making customer-specific organizational context—not generic acquired data—the valuable layer.
- Enterprise resistance is aimed at “Chinese models,” not open models generally, and Atlassian currently uses no Chinese models in production. It does use Nemotron, Mistral and other open models, routing each job toward the best price-performance outcome. Cannon-Brookes says US models increasingly use Chinese models to train themselves, making national labels less technically clean, but concedes customer concern is real today.
- The jury remains out on whether most specialized models can sustain an advantage over rapidly improving general models. Proprietary datasets and highly specialized applications such as robotics or world models might justify specialization, but many supposed proprietary models are merely billion-dollar foundations tailored for roughly $1 million. The falsifier is clear: they must maintain “exponential gains” in price-per-watt performance as the generic frontier advances.
- AI cost management is becoming a core enterprise-software capability, not a back-office exercise. Atlassian operates an internal AI gateway team for model routing, blending and harness selection across the quality-cost-speed triangle. Cannon-Brookes initially prioritizes magical adoption—“If five people are using a feature, I don’t care how much it costs”—then optimizes aggressively if usage reaches five million.
- Cannon-Brookes has revised downward his estimate of future inference prices, although lower unit costs will drive much more usage. Capital, talent, chips, data-center design and distillation should push price-performance down “at a precipitous rate,” but he rejects infinite demand: technology production already exceeds customers’ ability to absorb business change. A supercomputer-equipped phone still does not burn its CPU 100% of the time.
- As technology becomes abundant, competitive scarcity shifts from privileged knowledge to creative application. Borrowing Culture Amp founder Didier’s phrase, Cannon-Brookes says business is moving “from a world of information asymmetry to imagination asymmetry.” The personal counterweight is deliberate slowness: agents multiply activity and demands on attention, so leaders must preserve deep thinking and learn to “slow down to go fast.”
Deep dive
1. Agents will redistribute work without collapsing every interface
Harry Stebbings opens with the premise that agents could reduce software companies to data providers. Cannon-Brookes rejects the generalization: an agent can buy a charging cable when someone wants the “pedestrian, I just wanna get the thing done” outcome, while shopping involving research, taste or enjoyment still benefits from an experience.
Enterprise collaboration sits toward the creative, collaborative end of that spectrum. Agents will advance workflows and remove lower-value labor, letting people focus on creative work, but Cannon-Brookes does not view the resulting job changes as inherently negative—or bad for Atlassian.
His architecture allows horizontal agents, deep vertical agents and application interfaces to remain valuable simultaneously. Atlassian already connects through Claude, Codex, Cursor, Teams and Slack, yet Cannon-Brookes is “a big bear on the magic button thesis of AI”: no single uber-agent will do everything magically.
“Headless is brainless by definition” is not an argument against MCP or CLIs; Atlassian says it has one of the world’s top-five, perhaps top-three, MCP servers. It is an argument against abandoning the customer experience while relying on others to expose the value of your product.
His cross-product examples reinforce the point: Atlassian is adding Figma and Databricks functionality inside its platform, while Atlassian functionality can also be reached from Figma. A user could select a number in Confluence and retrieve an answer or small chart from Databricks or Snowflake without turning Confluence into a full BI tool. He calls this A2A across multi-craft teams; each interface retains its own strengths.
2. AI reopens foundational design problems
Cannon-Brookes believes design becomes more important as technology creation gets cheaper. An app better than 90% of alternatives today might beat only 20% in five years because expectations compound; continuously upgrading experience requires difficult, sustained craft rather than a one-off redesign.
The last comparable reset was mobile. Atlassian is now reconsidering how agents appear in collaboration software, how humans invoke them, how agents operate on people’s behalf, and how products communicate provenance, accountability and AI involvement without drowning users in implementation traces.
The primitive interface is inadequate: “Text boxes suck.” A blank line appears to offer unlimited power, yet most users treat it like Google Search. The open design question is how software reveals the right capabilities at the right moment as available agentic functions continuously change.
If Codex creates Harry’s pull request, should it appear as Harry’s work, Codex’s work or “Codex on behalf of Harry”? That distinction enables views containing “just the human comments” or “just the agentic comments.” Likewise, watermarking cannot distinguish a 3,000-word generated essay from a human essay lightly refined by AI: “same set of text, same watermark.”
3. The interface for supervising agentic work is still a design challenge
Cannon-Brookes finds the “software factory” analogy imperfect because software is creative rather than mechanistic. Its useful insight is that people increasingly improve and supervise the factory instead of manually stirring, chopping or stamping every object moving through it.
Humans still initiate work, steer it, co-create, handle exceptions and decide when to intervene. The design challenge is confidence: users should be alerted when something goes wrong, spared notifications when it goes right, and still able to inspect detail without confronting “a sea of things going on.”
Harry predicts that human-created work will require watermarking once content becomes 90% AI-generated. Cannon-Brookes offers the reverse possibility: within five years, society might neither know nor care what remained untouched by AI, just as audiences do not ask whether films were color-graded or used invisible CGI.
He preserves one boundary: entirely synthetic “AI-generated slop” differs from assisted craft. Yet the likely equilibrium is normalized synthetic production alongside greater differentiation for artisanal work and human judgment—not a simple good-versus-bad label attached to any AI-touched artifact.
4. Atlassian wants to index each company, not own its data
The Teamwork Graph combines decisions, projects, documents, people, code and services. “Data is stored in databases” is too narrow: Cannon-Brookes counters, “Nah, there’s a lot of data stored in PowerPoints,” including facts that never reached a structured system.
Its job resembles Google’s web index: organize and infer in advance so an agent can retrieve an answer cheaply at scale. Atlassian adds lexical and semantic search, an agent-oriented query language through MCP and CLI, and support for each customer’s custom entities—from stocks to materials, hospitals or beds.
Freshness matters more than static volume. A company’s social graph changes whenever employees, projects and working relationships change; current facts must therefore stay current, while history should persist as lessons and memory rather than obsolete detail.
Cannon-Brookes’s own example makes the distinction concrete: the Java code he wrote in 2003 is useless to Atlassian in 2026, but knowing he was once a programmer may help software explain something appropriately. Hence buying a company for its dataset offers no obvious missing piece: “It’s the customer’s context. It’s not our context.”
He distinguishes this from data marketplaces seeking large-scale, generic datasets to sell across providers. He cites the example—while the conversation leaves the buyer attribution uncertain—of Spirit Airlines data sold for $30 million, including 1 million emails. That kind of generic dataset may serve broad intelligence platforms, whereas Atlassian is assembling a live, customer-specific graph.
5. Enterprise resistance targets Chinese models, not open models
Cannon-Brookes confirms “some customer pushback in some places” to “Chinese models,” while saying there is no comparable rejection of open models. Atlassian uses Nemotron through its NVIDIA partnership, some Mistral and other open models, but “we don’t use any Chinese models in production.”
Atlassian’s job is application rather than ideological model selection: blend the best model for each task, price and desired outcome. The complication is provenance—Cannon-Brookes says US models increasingly use Chinese models to train themselves, creating feedback loops that make national labels less technically clean than customers may assume.
His Linux analogy: enterprises accept open-source software despite what he estimates may be roughly 40% of contributions coming from China, partly because the code is inspectable. He argues that AI may make such inspection even easier.
Harry’s pushback—worth keeping—is that technology and geopolitics cannot be separated when model builders and their families face state constraints. Cannon-Brookes declines the social-policy debate but distinguishes strategic control by America or China from whether an enterprise should use inspectable technology that is cheaper and performs better.
6. Specialized models must outrun a moving foundation frontier
“The jury is really out on the benefits of building specialized models.” Cannon-Brookes sees legitimate cases involving proprietary data, narrow applications, robotics and world models, but doubts that most specialized efforts can defend today’s price-performance advantage as general models improve.
He objects to companies being “fast and loose with words.” Atlassian fine-tunes specialized open-weight models but does not call them its own; spending roughly $1 million to tailor somebody else’s billion-dollar foundation is not equivalent to training from raw data.
Many branded corporate models, he argues, are generally Chinese foundations with domain intelligence added and a new name attached. Fine-tuning can be valuable, but describing the result as a wholly proprietary model obscures both its economics and its technical lineage.
His test is sustained, exponential price-per-watt gains over the latest foundation model. Agentic harnesses can repeatedly apply prompts and configurations—and better harness design matters enormously—but “it’s not model progress.” Many model-building startups from two or three years ago have already been “smashed” by faster generic-model improvement.
7. Cost discipline becomes product capability at scale
Harry poses the private-company temptation directly: if abundant venture funding buys a product that is 2%-3% better, why care about inference cost? Cannon-Brookes’s answer is that funding is unlikely to last forever, while enterprises are already intensely sensitive to price-performance as token spending becomes a larger budget.
Atlassian treats model routing, blending and harness selection as an internal capability. Its substantial AI gateway team continuously optimizes the triangle of answer quality, price and speed; smaller or less technical companies can sensibly outsource routing, but companies at Atlassian’s scale must develop the skill themselves.
Product teams are told, “Don’t over-optimize for cost in the short term. Build amazing, magical features.” If five people adopt one, cost is irrelevant; if five million do, optimization becomes urgent, with enough time between those points to change model choices or improve the harness.
Cannon-Brookes has changed his mind most on “the likely future price of inference”—it should fall further than he previously believed. Chips, data centers, distillation, capital and extraordinary talent are all attacking price per watt, making technological cost management a defining strength of great software companies.
8. Abundant inference shifts scarcity from information to imagination
Harry invokes Jevons paradox: cheap inference may simply produce vastly more inference. Cannon-Brookes agrees usage will explode but rejects infinity; “the production of technology is faster than the consumption of technology,” and customers already cannot absorb features quickly enough to transform their businesses.
The larger opportunity is a “creativity-driven economy.” Borrowing a phrase he credits to Culture Amp founder Didier, Cannon-Brookes says value is moving “from a world of information asymmetry to imagination asymmetry”: advantage comes less from knowing a hidden fact than from applying AI to problems in an original way.
His own shortfall is not imagination but protected time. Agentic work encourages “push the agent, do the thing” cycles while notifications from tools keep attacking attention. After a hectic year, he has spent the last two or three months deliberately recovering even a couple of hours weekly for long, year-to-year thinking.
9. Durable software and founders both require continual renewal
OpenAI and Anthropic are deep Atlassian partners but also operate at some level in applications. Cannon-Brookes calls the eventual boundary “a highly fought-over variable in the future,” yet argues that raw document creation is already commoditized without eliminating differentiated applications.
Confluence supports the case: its last disclosed scale was north of $1.5 billion in ARR, growing well above 25% annually, despite users having perhaps 20 other ways to write text. The differentiation lies in design, purpose and workflow integration—not the mere ability to place words on a screen.
Harry challenges him with the perception that Linear’s craft is more palpable than Jira’s. Cannon-Brookes calls that unfair, cites accelerating Jira and Teamwork Collection growth alongside foundational gains in design, speed and efficacy, and says Atlassian must “run its own race” across decades.
Asked about his comparatively long tenure, Cannon-Brookes answers “24.” His endurance comes from enjoying hard problems and smart colleagues, not simply money or status. He also links founder endurance to energy: early-stage founders carry livelihoods, fundraising, revenue, customers and reputation, and money or the right circles rarely sustain five or ten years of work on their own.
He describes wealth as a wonderful thing and a burden: a temporary responsibility to pursue “societally profitable,” rather than merely personally profitable, uses. His closing counsel joins humility with restraint: “wage peace on yourself,” create moments of slow thought and “slow down to go fast.”