Framer's Harness 2 Makes Its Agent Faster by Changing the Plumbing, Not the Model

Framer's Harness 2 write-up is one of the more useful agent posts this month, because the gains come from plumbing rather than a new model.
The change is eager execution. Instead of waiting for the model to finish writing a program and then running it, Framer runs each complete statement as it streams in, inside a Web Worker holding a copy of the document. Reads and dependent writes, using calls like tree.get and tree.patch, happen in one program instead of a back and forth of tool calls.
Framer reports 39 percent fewer cached input tokens per message on average, and says it is up to 33 percent cheaper and 20 percent faster with no measured drop in quality. Its own tool step figures are less tidy, quoted as both 2x fewer and 43 percent fewer, and there is no methodology, so read the numbers as direction rather than proof.
The takeaway for anyone choosing agent tools: ask how many round trips a task takes, not just which model sits underneath. External agents such as Claude Code and Cursor now get the same single-program path through Framer's MCP.