Discussion

tldraw gives coding agents a canvas, and their behaviour changes with it

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#3223

tldraw’s product engineering team has built a spatial harness that lets coding agents work inside a 2D canvas rather than only through a linear chat window. The important point is that the interface is not just decoration: the team says changing the environment changed how the agents behaved.

Watch Desk analysis

What happened

In a post published by tldraw on 22 September, the product engineering team describes experiments that moved from single-shot shape editing to multi-agent browser simulations and then to local coding agents writing code directly into the canvas editor. The work treats the canvas as an environment an agent can inspect and manipulate, not simply as a place to display its final answer.

The team’s account says the spatial harness was designed to help coding agents understand and interact with a 2D canvas. That distinction matters. An agent working in a canvas has to deal with position, relationships between objects and the visible state of a shared workspace, rather than shuffling through a tidy sequence of text instructions. The source describes this as changing agent behaviour compared with traditional linear chat interfaces.

Why it matters

Most coding-agent demos still imagine software as files, terminals and messages. A canvas points towards a broader class of tools in which an agent has to operate inside a visual, stateful environment. That could include diagramming, design, planning and collaborative whiteboards, where success means changing the workspace correctly rather than merely producing plausible prose or code.

The practical lesson is modest but useful: agent capability may depend as much on the environment and feedback loop as on the underlying model. Give the same system a different workspace and its strengths, weaknesses and mistakes may become different beasts entirely. Software interfaces, it turns out, can have opinions.

Our read

tldraw’s account is a product-engineering report, not an independent benchmark, and it does not establish that canvas-based agents are broadly more capable. It does establish a worthwhile direction for testing: compare agents in realistic spatial environments, measure what they actually complete and examine how the interface shapes their decisions.

The interesting question is no longer only whether an agent can write code. It is whether the workspace gives the agent enough structure to understand what the code is meant to change.

What to watch

  • Task completion:
    whether agents can reliably create, edit and repair canvas objects rather than merely describe them.
  • Spatial reasoning:
    how well they handle layout, relationships and changes made by other agents or people.
  • Evaluation methods:
    whether future tests compare canvas and chat workflows under equivalent tasks.
  • Tool adoption:
    whether other editors and collaborative applications build similar spatial harnesses.

Discussion spark: Should coding agents be judged mainly by the model they use, or by whether their interface gives them a workable environment to reason and act in?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.