Code mode lost its first race. Then we gave it muscle memory
Blog post from Blacksmith
Blacksmith introduces [code]smith, a general-purpose coding agent designed to narrow the gap between local development, CI environments, and agent execution by using the same underlying components as CI. Its “code mode” lets agents search for MCP tools, interact with their TypeScript-style interfaces, and execute multi-step tasks in a sandbox, reducing the amount of raw tool data that enters the model’s context compared with conventional tool-calling loops. In a test involving Linear issues and comments, cold code mode initially performed worse than classic tool calls because MCP servers generally expose input schemas but not the shapes of returned data, forcing the agent to make incorrect guesses and run extra scripts. Although slower in that first run, code mode still used less context and cost slightly less. Blacksmith addressed this limitation with “Muscle Memory,” a cache that records inferred response-field names and broad types without retaining values, allowing future sessions to understand tool outputs immediately. In a repeat test with cached shapes, code mode completed the task in 56.6 seconds, 27% faster than classic calls, while using the smallest context and roughly half the cost.
No tracked trend matches for this post yet.
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.