fx drops interactive subagents, smaller binary

While terminal agents like Claude Code and Cursor dominate mindshare, fx demonstrates that tiny native binaries with zero runtime dependencies and aggressive prompt-cache optimization deliver a faster, cheaper feedback loop. Splitting workflows between frontier planning models and cheap execution subagents is rapidly solidifying into the standard architecture for autonomous coding.

• –Subagent delegation operates directly within the terminal scrollback without taking over the full screen, letting developers switch between running agents using Ctrl+X or arrow keys.
• –Hierarchical multi-agent setups require just a single instruction in ~/.fx/AGENTS.md, routing high-level planning to flagship models and execution to lightweight models.
• –A 48% reduction in tool definition payload combined with an 82% prompt cache hit rate dramatically cuts API token costs and latency during long debugging sessions.
• –Packaging a complete coding agent into a 6.3MB Zig binary removes Node.js and Python runtime dependencies, making it effortless to run in sandboxes and CI runners.
• –Built-in support for Vercel Token Pass, custom OpenAI-compatible endpoints, and local LLMs avoids vendor lock-in to any single model provider.