Figma's Matt Colyer on AI Agents and the Future of SaaS
the gist
Matt Colyer argues that the 'SaaSpocalypse' is actually a goldmine for software, as AI agents move beyond text-box interfaces to handle complex, divergent design workflows and cross-tool automation.
The 'SaaSpocalypse' as a Growth Engine
Matt Colyer rejects the narrative that AI will render SaaS obsolete. Instead, he views the current era as a massive expansion in the number of software builders—potentially growing from 30 million to over a billion. He argues that while AI makes it easier to prototype, the operational burden of maintaining software (e.g., managing infrastructure, SMTP, and reliability) remains a significant friction point. Consequently, he finds himself buying more SaaS tools than ever to outsource the maintenance of his own custom-built agents.
Moving Beyond the Text Box
Colyer emphasizes that current AI interfaces, which rely heavily on linear chat, are ill-suited for design. He advocates for a 'diamond-shaped' design process: a divergent phase to generate numerous options, followed by a convergent phase to refine the best ideas. Chat interfaces struggle with this because they are inherently linear. Figma’s strategy involves moving agents directly onto the infinite canvas, allowing users to manipulate visual frames, iterate on layouts, and cluster concepts rather than just prompting a text box.
Closing the Loop with MCP
Figma is leveraging the Model Context Protocol (MCP) to bridge the gap between code and design. By allowing external agents to interact directly with the Figma canvas, developers can pull live UI into a design environment, make adjustments, and push changes back. This removes the 'drudgery' of manual UI updates and allows for a more fluid transition between engineering and design workflows.
The Review Bottleneck
As AI agents become more capable of generating code and design, the primary bottleneck in product development has shifted from 'creation' to 'review.' Colyer notes that while agents are excellent at executing tasks, they often struggle with nuance, leading to over-correction or literal interpretations of feedback. The future of product work, he suggests, lies in building better systems for human-in-the-loop oversight and context-aware agent personalization.