According to a recent LinkedIn post from Blacksmith, the company is experimenting with a “code mode” approach intended to accelerate how AI agents call MCP tools. The test compared this mode with classic tool calling on the same Linear task and initially found code mode to be slower, despite its design for fewer round trips and smaller context.
The post suggests that the performance gap stemmed from how most MCP servers only ship input schemas, forcing agents to guess response structures and incur extra turns when wrong. In response, Blacksmith reports developing a “shape cache” that stores tool output structures across sessions, with a subsequent rematch showing code mode 27% faster, 47% cheaper, and using 42% less context than classic calling.
For investors, this update indicates ongoing optimization of Blacksmith’s agent tooling infrastructure, potentially improving cost efficiency and response speed for customers. If these gains translate into more competitive AI workflows and better unit economics, the company could strengthen its value proposition in the developer and enterprise AI tooling market, supporting long‑term scalability and margin improvement.