Skip to main content

​
Software is Being Rewritten — by Agents

We are entering a world where humans won’t be writing most software. AI agents will. And agents will be creating other agents. When Claude Code scaffolds a project, installs dependencies, writes tests, and iterates on failures — all in a single session — the bottleneck isn’t the model. It’s the runtime. In December 2025, Anthropic made their first-ever acquisition: they bought Bun. Not a model company. Not a data company. A JavaScript runtime. That tells you everything about where this is going. We’re making the same bet.

​
The Runtime is the Agent’s Operating System

Every tool an AI agent calls is a runtime operation — HTTP requests, file I/O, code execution, JSON parsing. An agent making dozens of tool calls per session doesn’t experience latency the way a human does. For a human, 500ms is imperceptible. For an agent in a tight loop, it’s the difference between fast iteration and a stall. Node.js was designed for a world where humans start a server and it runs for hours. Bun was designed for a world where processes spin up, execute, and terminate in milliseconds. That’s the world agents live in — generate code, run it, check the output, fix the errors, repeat. The runtime needs to keep up. Anthropic understood this before anyone else. Claude Code ships as a Bun executable. So does FactoryAI. So does OpenCode. The pattern is clear: the runtime is becoming the agent’s operating system.

​
One Binary, Zero Configuration

The Node.js ecosystem asks you to choose: which package manager, which bundler, which test runner, which TypeScript compiler. Five tools, five configurations, five places where things break. Bun is one binary that does all of it: 
bun install        # Package manager
bun run dev        # Runtime
bun run build      # Bundler
bun test           # Test runner
This isn’t just developer convenience. For an AI agent that needs to spin up projects, test, bundle, run, and iterate at speed, that fragmentation is friction. Every tool choice is a decision point. Every configuration file is a place where an automated workflow can break. Bun eliminates the entire category of problem. Jarred Sumner, Bun’s creator, said it directly: Bun’s job is to be the best place to build, run, and test AI-driven software.

​
TypeScript Without the Ceremony

Bun runs TypeScript natively. No compilation step. No tsconfig.json. No build pipeline between writing code and running it. aixyz packages ship raw .ts files and Bun handles them directly. When an agent generates TypeScript, it can execute it immediately — the generate-run-verify loop has zero gap. This is the kind of simplicity that makes agent-driven development practical at scale.

​
Compile Once, Run Anywhere

Bun compiles projects into single-file executables. No runtime installation required. No dependency management on the target machine. This is how Claude Code distributes to millions of developers — one binary, every platform. For aixyz agents deploying to serverless platforms like Vercel, Bun’s cold starts are fast enough that your agent responds quickly even on first invocation: 
{
  "buildCommand": "bun run build",
  "bunVersion": "1.x"
}

​
Built-in Testing

Bun includes a Jest-compatible test runner with near-instant startup. aixyz uses bun:test for agent testing: 
import { describe, expect, test } from "bun:test";
import { ToolLoopAgent } from "ai";
import agent from "./agent";

test("default export is a ToolLoopAgent", () => {
  expect(agent).toBeInstanceOf(ToolLoopAgent);
});