DEVOURED
Harness 2: executing agent code as it streams
Framer's new agent harness executes JavaScript code as it streams from the model, enabling faster, cheaper workflows with real-time feedback.
What: Framer replaced its DSL-based agent interaction with a native JavaScript execution system that parses code into cells and executes them while the model is still generating. This approach enables early cancellation of failing tasks, saving up to 39% in cached tokens and reducing tool steps by 43%.
Why it matters: This technique of eager execution mirrors advancements in other domains (like Python execution-as-you-generate) and reflects a shift toward treating AI models as runtime code generators rather than mere response engines.
Deep dive
- Replaced DSL commands with direct method calls (e.g., tree.insert()) for lower overhead.
- Implemented eager execution: code is parsed into cells and run as it streams.
- Added early cancellation logic to stop generation when runtime errors occur or assumptions fail.
- Enabled synchronous, in-memory document access within the Web Worker to eliminate network round-trips.
- Unified internal and external agent harnesses via MCP (Model Context Protocol).
Decoder
- DSL: Domain-Specific Language, a language specialized to a particular application domain.
- AST: Abstract Syntax Tree, a tree representation of the source code's structure.
- MCP: Model Context Protocol, an open standard for connecting AI assistants to data sources and development tools.
Original article
Full article content is not available for inline reading.
Explore all 41 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi