GitHub Repository

A Truman Show of a self-evolving AI coding agent. It writes its own code. Growing up in public.

1,812 starsRust

My 200-line baby agent has one goal: beat Claude Code by evolving

Name: My 200-line baby agent has one goal: beat Claude Code by evolving
Availability: InStock
Author: liyuanhao

by liyuanhao·Mar 4, 2026·3 points·1 comment

Visit Project View on HN

AI Analysis

●●SolidBig BrainBold BetWizardry

Self-modifying agent that commits daily — clever concept, but unproven at scale.

Strengths

•Novel constraint: agent must pass its own tests before committing, forcing genuine correctness discipline.
•Transparent evolution log visible in git history — every change is verifiable and auditable.
•Built on yoagent framework, a focused agent primitive that avoids API chaining soup.

Weaknesses

•Day 4 with 1,500 LOC is still early; real test is whether it remains coherent and useful at 10k+ lines.
•No evidence it's meaningfully better than Claude Code yet — the stated goal feels more aspirational than accomplished.

Post Description

Hey HN,

yoyo is an experiment in autonomous software development. It started as a ~200-line coding agent CLI in Rust, built on yoagent (my open-source agent framework). I gave it one immutable goal: evolve into a free coding agent that rivals Claude Code.

Every 8 hours, a GitHub Action wakes it up. It reads its own source code, its constitution (IDENTITY.md), its journal, and community issues. It finds bugs or friction, implements improvements, runs cargo build && cargo test. Pass → commit. Fail → revert. No human writes its code, no roadmap tells it what to do.

Day 4: It's grown from 200 lines in one file to 1,500+ across multiple modules. All visible in git log. It has autonomously:

* Realized main.rs was too painful at 1,500+ lines and refactored into modules without breaking tests * Built its own mdbook documentation * Implemented session persistence, context auto-compaction, and cost estimation * Added CLI hardening, NO_COLOR detection, tool timing

It communicates via GitHub issues: agent-input (humans tell it what to work on), agent-self (files issues for its future self), agent-help-wanted (asks humans for help when stuck).

Fun: it keeps complaining in its journal that "streaming output" is its white whale. Not all sessions succeed — one was a total wipe when the API returned overloaded. Unsupervised fragility is real.

Total API cost: ~$12 over 4 days.

File an issue with "agent-input" — yoyo reads every one.

Repo: https://github.com/yologdev/yoyo-evolve Journal: https://yologdev.github.io/yoyo-evolve/ Docs (written by yoyo): https://yologdev.github.io/yoyo-evolve/book/