November 23, 2026CodingOpen SourceAgents

fast-anything Lets the Agent Write Rust, Then Makes Production Decide

Rewrite it in Rust has always been a slogan with a bill attached: you port the whole thing, break something subtle, and spend a month finding out what. fast-anything takes the opposite bet. Port only the few functions that actually eat your CPU, let a coding agent write them, and refuse to switch anything on until the new code reproduces what the old code did on real traffic. The repo was created on Saturday, shipped v0.7.0 on Sunday, and sat at 172 points on Hacker News with 1,776 stars by the end of the day. https://github.com/EL4CTEO/fast-anything

The loop is three commands. scan profiles your program and ranks functions by time spent times how portable they look. port records up to 500 real calls to the chosen function, generates a one-function Rust crate plus a TASK.md, and hands it to an agent: Claude Code, Codex, or anything that reads stdin. verify replays every recorded call plus derived edge cases, NaN, plus and minus infinity, negative zero, non-ASCII strings, empty inputs, inputs up to 16 times bigger, shrunk slices. Floats must match within 1e-9, errors must happen where the original errored, and a port that is slower per call is rejected too. Nothing is enabled until every case matches.

The README's best example is a failure. One port compared bytes instead of characters, the classic Unicode bug a human reviewer skims right past. 148 of the 500 recorded real calls caught it. That is the whole argument in one number: the agent's confidence is worth nothing, the recorded behavior of your own production code is the spec.

The speedups are self-reported, measured on the author's i5-10500, and should be read that way: a Python fuzzy dedupe from 4.91 seconds to 42 ms (116x), PHP 55x, Go 21x, TypeScript 20x, Ruby 14x, C# 13x, Java 3x, C 3.4x, C++ 2.7x. There is a sharp lesson buried in there too. Porting only the inner levenshtein call gave 1.9x where porting the outer loop gave 35x, because the boundary crossing costs more than the math. It covers 11 languages and ships a Claude Code plugin and an MCP server with 8 tools. MIT license.

Two days old, so treat it as a sharp idea rather than proven infrastructure. But it is the cleanest version yet of a pattern that keeps winning: let the model write the code, and let something outside the model decide whether it ships.
← Previous
OpenRouter Pulls 125 Endpoints From Zero Data Retention. The Label Was Never Tested
Next β†’
Worker 755 Makes You the Subagent, and the AI Writes Your Review
← Back to all articles

Comments

Loading...
>_