The language holding our agents back.

Theo (t3.gg) argues bash is only a stepping stone for AI agents and TypeScript execution environments are the real next layer.

• Cloudflare’s TypeScript-SDK approach cut average tokens per response from 43,500 to 27,000 (~40%) and lifted benchmark accuracy from 25.6 to 28.5.
• One MCP server example consumed 72,000 tokens — roughly 40% of context — just for tool descriptions, before any actual work.
• Whole-codebase context dumps (e.g. Repomix) cost T3 Chat over $100,000 in wasted API spend from users pasting repos as cheap Cursor alternatives.
• Permission-approval fatigue is so severe that most users, including Theo, run agents in full skip-permissions mode, eliminating the safety layer entirely.
• Google optimized models for large-context retrieval; OpenAI and Anthropic focused on tool-calling for targeted fetch — Theo cites this as a core reason Google models underperform on agentic coding tasks.
• TypeScript V8 isolates allow hundreds of concurrent agent sessions on one Linux kernel with no Docker overhead and no cross-user file access.
• Vercel built a virtual bash written in TypeScript that never touches the real filesystem; Rhys Sullivan built Executor as a TypeScript-native execution environment for agents.
• Giving agents one bash tool outperforms giving many specialized tools because each additional tool bloats context and increases non-determinism.

2026-04-07 · Watch on YouTube