Philipp Schmid suggests that Agent tool invocation is moving toward Code Mode

On November 2, 2026, Philipp Schmid put forward the judgment that Agent tool invocation is moving from a small number of tools and dynamic tool search toward folding everything into Code Mode. He wrote an article explaining how Code Mode works, why a typed JS/Python sandbox can outperform bare bash, and why MCP will make a strong comeback in 2027, and he included a minimal example of Gemini 4 Argon. The next day, a report further summarized his views: the model receives only a single exec(code), and within a TypeScript/Python sandbox it strings together queries, filtering, modifications, and parallel operations, with most intermediate results not entering the context; in implementation, the tool catalog is compiled into compact type declarations, text()/image()/audio() control which outputs enter the next round, and store/load keep state in host memory. The report also stated that Pi, Factory Droid, OpenAI Codex, and others already default to or support Code Mode, and that after Factory switched, Terminal-Bench 4 accuracy was about +3.5 and model cost was about -15%.

AI-generated from public reports

1. Nov 2· 2 reports

   Philipp Schmid proposes that Agent tool calling is moving toward Code Mode

   Source:Shao Meng