Framer’s AI design agent now starts changing your website before the model has finished writing the instructions for the change. Framer, the website builder, described the rebuild in a post on its blog on 2 November 2026. It calls the new version Harness 2 and says the agent costs up to 33% less and runs 20% faster.
A harness, in plain words, is the software wrapped around an AI model that lets it act. It feeds the model information, takes what the model produces, and turns that into real edits. Framer’s first harness had the model emit short commands, one per change, so a page visibly came together as users watched. Bulk work was clumsy. Updating a property across 100 frames took 100 commands, and any job that read a value before writing it needed separate trips back to the model, each one replaying the earlier context.
The new version has the model write a JavaScript program instead. Framer’s example is a request to find every use of an old brand color and swap it out. That can now be one program with loops and conditions: it inspects the document, decides which nodes to change, and applies the edits, with no call back to the model at each decision point.
The timing is the clever part. As code arrives, Framer parses it, groups complete statements into chunks it calls cells, and runs each cell in a background worker straight away. The first edits appear on the canvas as the model keeps writing the remainder. Comments in the code can double as progress messages, so a long job tells the user what it is doing.
Running code mid-stream also lets Framer abandon a doomed attempt. The agent is prompted to check its assumptions ahead of any step that relies on them, such as whether a node exists. If a check fails, or the program returns early because nothing needs changing, Framer cancels the remaining generation. Edits already applied stay put. Cancelling saves writing time and the output tokens a provider bills for, though Framer notes the exact saving depends on how far the model had got. The error message also explains what went wrong, so the next attempt can fix the bad assumption.
Several supporting changes make that speed possible. The agent works on a copy of the document held in memory inside the worker, so reads and writes happen instantly instead of waiting on round trips. Recoverable write problems no longer kill a run: the system applies valid properties, clamps out-of-range values, skips unsupported ones, and lists the warnings at the end.
Outside agents get the same treatment. When Framer for Agents launched in June, the in-app agent and tools such as ChatGPT Desktop, Claude Code, and Cursor worked differently. Now an outside agent can submit one program through MCP, the protocol that connects AI assistants to apps, and it runs inside Framer with direct access to the document.
On average, Framer reports, a message now needs 43% fewer tool steps, plus 39% less cached input in tokens, and quality showed no measured drop. The post cites no independent test, and the gains come from fewer model round trips, so they should be largest on bulk jobs like site-wide restyles and smallest on one-off tweaks.
Framer also disclosed a prior claim to the idea. It says it believed it was first, then found a paper titled “Executing as You Generate”, first published on 1 April 2026. Its authors, Zhensu Sun and co-authors from Beihang University and Singapore Management University, built a system that applies the same approach to Python. Framer says the paper includes one optimisation it missed, clearing finished statements from the parse buffer, and plans to adopt it.
Any team building an agent that edits a live document now has a public pattern to copy: stream the code, run it as it lands, and cancel the moment a precondition fails.
Reported by Framer on its blog on 2 November 2026.
