20VC: Atlassian CEO on Why Agents Won't Replace Interfaces | Why Enterprises Won't Trust Chinese Models | Why The Jury Is Still Out On The Benefits of Specialised Models with Mike Cannon-Brookes

Atlassian's CEO bets design and data beat owning models

Atlassian's CEO argues the AI era rewards design and company context more than owning models. Agents will take over functional tasks, but collaboration and taste still need interfaces, so Atlassian is building for people and agents to work together. He expects generic models to keep improving and inference to get cheaper, which leaves the data graph and the experience around it as the durable position.

Takeaways

• Agents will not replace the user interface. Purely functional tasks like buying a charging cable will go to agents. Work involving taste, interaction and collaboration will not. Mike Cannon-Brookes calls giving up on the interface a brainless bet and says design will differentiate in the AI era.
• Provenance is the new foundational design problem. When humans and agents produce content together, it is unclear who did what, as with a Codex pull request. Mike expects people to stop judging AI involvement, and expects human-made artisanal work to stand out more.
• Specialised models remain unproven against generic ones. Mike says the jury is out. They can win with proprietary data or robotics, but many startups that built their own models were overtaken by generic ones. Agentic harnesses add value, though he does not count that as model progress.
• Inference costs will fall further than expected. This is Mike's biggest change of mind in twelve months. Cheaper inference drives more usage, and technology production currently runs ahead of consumption. Atlassian tracks customer token spend and routes models through an AI gateway team, but teams should not over-optimise cost early.
• Enterprises resist Chinese models, not open ones. Atlassian is customer led and uses NVIDIA's Nemotron and some Mistral models, with no Chinese models in production. Mike sees customer pushback on some Chinese models but none on open models, and says many claimed in-house models are fine-tunes of open weights.
• Company context is Atlassian's data advantage. The Teamwork Graph works like Google's web index, built in advance for each company. Agents and harnesses read from it and write to it through MCP. Mike says companies will not sell internal data, and Atlassian will not copy the $30m Spirit Airlines data sale.
• Labs are partners while the boundary stays blurry. OpenAI and Anthropic are deep Atlassian partners, yet the line between labs and application companies is uncertain. Mike argues design, integration and purpose matter more than writing content, citing Confluence at over $1.5bn in ARR growing above 25%.

Where Mike and Harry disagree

• Harry floats horizontal agents replacing vertical software. Mike expects both to exist, along with agents that run on platforms such as Claude, Codex, Cursor, Teams and Slack.