"Code mode is the future": LLMs should write programs instead of calling many tools
ayirpelle · x · 2026-11-08
In an X thread, springrod argues that "code mode" is the future: rather than calling many individual tools step by step, LLMs should write programs.
The claimed benefits: it saves context window, plays to models' pretraining on huge amounts of code, and a program is a better way to expose business understanding than a pile of tool schemas. He follows up by asking what language LLMs should code in. This stands in clear contrast to the per-tool-call pattern popularized by MCP, and is a key fault line in current agent architecture debates.
Related event: 'Code Mode' Debate: Should LLMs Write Programs Instead of Calling Tools(4 posts)→
More from coding & agent
- Dev Uses AI Agent to Branch, Update and Wire Two PowerShell Modules in One Pass — dfinke · 2026-11-08
- mitsuhiko: djs' JS "heap dumps" may beat real dumps for durable coding agents — mitsuhiko · 2026-11-08
- Claude Code builds 1M-zombie GPU crowd simulation in one session, 12s clip in 21 min — ravisparikh · 2026-11-08
- 2026 AI office agent landscape: 150 companies, 16 layers, model labs selling straight to enterprises — FinanceYF5 · 2026-11-08
- exedev quietly launches per-minute sandbox VMs via 'ssh exedev new --sandbox' — davidcrawshaw · 2026-11-08
- Vibe-designed programming languages flood in, and this PL veteran is not happy — zetalyrae · 2026-11-08