Max Drake

Max Drake is a product engineer at tldraw, building spatial interfaces that help people create with software and collaborate with AI agents. Drake’s work connects tools for thought, canvas-based model interactions, and animation through a recurring concern: how movement, arrangement, and visible activity can make powerful software easier to direct.

From mechanical engineering to tools for thought

Drake studied mechanical engineering at MIT. Earlier research and design work, including work at Fermat, preceded the role at tldraw and forms part of a career spanning engineering, interface design, and tools for thinking. Drake’s canvas experiments with language models began before ChatGPT’s release, developing into work on agents that can interpret drawings, navigate a workspace, and collaborate with people.

Drake’s essay on LLMs and spatial canvases articulates an important connection between those interests. A canvas can do more than display a model’s output: the arrangement of information can help specify an operation. Moving a tool or grouping information can trigger model operations scoped to the relevant part of the workspace. Spatial interaction becomes a way to express what the model should work on, allowing people to act directly on the material instead of describing every relationship in a text prompt.

At tldraw, Drake works within an infinite-canvas environment where people can manipulate shared objects and see collaborators’ cursors, selections, and viewports. Drake’s agent work extends that collaborative setting to software participants. The challenge has two parts: models need help understanding and changing two-dimensional space, while people need interfaces that reveal what agents are doing and how their work fits together.

Making spatial work understandable

• Spatially scoped model operations. Drake’s early canvas thinking treats placement and grouping as meaningful inputs. A tool’s position and the information gathered around it can determine the scope of a model operation. This gives spatial organization an active role in computation, rather than leaving it as decoration around a chat interface.

• Canvas understanding requires an engineered harness. The tldraw team’s Teach experiment combined screenshots, structured canvas data, and instructions for interpreting objects and making edits. A request to make a drawn mouse blow out a candle required the model to recognize an arrangement of ordinary shapes and place wind and smoke appropriately. The agent starter kit extended this into continued action: setting goals, changing the viewport, searching elsewhere on the canvas, and editing what it found. Drake’s spatial-agent work connects visual interpretation to the navigation and action loop needed to complete a task.

• Visible agent states are functional interface design. The team’s Fairies experiment represented agents as distinguishable animated characters sharing a multiplayer canvas. Appearance helped people recognize individual agents; animation communicated activity such as working or waiting. An orchestrating agent could assign work, wait for completion, and receive a notification to review the result. Drake’s argument is practical: with several agents operating at once, people should be able to understand their activity without reading every message.

• Executable dependency graphs. Drake’s Tech Tree prototype grew out of the large dependency boards the tldraw team used around product launches. Those boards made prerequisites and progress visible; Drake’s prototype connected the tasks to coding agents that could perform the work. People could launch an agent, inspect progress, open a pull request, or turn a drawing into context for an assignment. Multiplayer support let colleagues add and edit tasks in the same workspace. The canvas became both a shared project overview and an interface for directing execution outside it.

These experiments also exposed a constraint: agents built entirely around a canvas-specific harness can struggle to reach tools and activity beyond that environment. A desktop approach exposed the running tldraw editor through a server so an external coding agent could manipulate it with JavaScript while retaining access to local files and other tools. This opened a path between spatial interfaces and work performed elsewhere, although Drake’s live specification-to-demo request remained unfinished when the recording ended.

Animation through performance

More recently, Drake co-developed tldraw flash with Audrey Leo and Leo Knudsen. Its performance-based animation approach lets people record movement by manipulating objects, then edit that recorded movement. Instead of requiring users to specify every motion in advance, the interaction captures a performance and leaves it available for refinement.

Flash extends Drake’s interest in making direct manipulation computationally useful. In Drake’s spatial-model essay, arranging material helps define an operation; in the agent interfaces, position and animation reveal coordination; in Flash, performed movement becomes editable creative material. Across these projects, the workspace gives people concrete ways to express intent and continue shaping the result.