In early October 2026 I shared a series of LinkedIn posts about building agentic software factories. I explained how naïve tool‑calling bloats prompts, and why Pi’s Code Mode lets LLMs generate sandboxed TypeScript/JavaScript that calls MCP tools via a programming API, returning only the final result. I also announced Hugo Bowle‑Anderson’s two‑week live course (starting Oct 13) with a promo code for a discount and Modal credits, described the Decode 19‑task benchmark results, the error‑analysis pipeline using Kitaru, and the Jev zero‑shot classifier for cheap decision routing.
I adopt Code Mode because “Instead of exposing every MCP tool directly to the LLM, you expose them as a programming API.”
I shift my engineering focus to meta‑engineering: “You build the factory that builds the software.”
On the Decode benchmark, Qwen3.6‑35B achieved “Qwen3.6-35B → 95% pass@1”.
I set up an error‑analysis pipeline: “Kitaru by ZenML records Decode runs as sessions, helps turn recurring failures into cohorts, and lets me replay those same sessions after changing the harness.”
I integrate Jev for cheap routing: “Take the user's prompt and decide which model should handle it.”