autor3search
Point your coding agent at your repo. Come back to faster code.
FreeAI CodingOpen Source4 impressionslaunches Oct 271 comment
Comments
>log in to commentHey DevHunt 👋 I built autor3search because every "AI optimized my code" demo I saw skipped the hard part: measuring it. So the design starts from an adversarial premise. Assume the agent will eventually try to win by editing a test or a benchmark, and make that impossible. It's early but working, and every number we publish was measured on a real library. If you have a repo with benchmarks, I'd love for you to run it and tell me what it found, or what it didn't. Feedback very welcome!
AI coding agents are great at claiming they made your code faster. autor3search makes them prove it.
Your agent proposes one optimization at a time. A harness it can't edit checks the change against your frozen test suite, benchmarks it against a pinned baseline with an interleaved A/B design, and returns a verdict: KEEP or DISCARD.
- 🔒 Cheat-proof by design: tests are restored from a golden copy before every run, and the metric lives outside the repo where the agent can't reach it. Noise is discarded, regressions are rejected.
- 📜 Fully auditable: each accepted change lands as its own commit, and every experiment, failures included, is logged in results.tsv.
- ⏹️ Stop anytime: nothing already accepted is thrown away.
- 🌍 8 languages, same rules: Go, Python, TypeScript, JavaScript, Java, C#, Rust, Swift.
Real results, not illustrations:
- go-humanize: −40.2% geomean (Bytes 124 → 37 ns/op, allocations 4 → 1)
- apple/swift-asn1: 10.98× (local clone, not upstreamed)
- Python humanize: 5.81%, no regressions
Setup is one prompt: paste it into Claude Code, Cursor or any coding agent from inside your repo. MIT licensed. Inspired by karpathy/autoresearch.
- Autonomous AI‑driven code optimization
- Frozen measurement harness isolates agent from source
- One commit per accepted change
- Experiments run until stopped, preserving all data
- Real case studies show up to 57% speed gains
- Supports Go, Rust, Java, JavaScript, Python
autor3search runs AI coding agents against a frozen test suite to benchmark and verify code optimizations.
- for
- Developers who want measurable, auditable performance improvements from AI agents.
- pricing
- open source
- license
- MIT
Key features
- Cheat-proof harness — Tests are restored from a golden copy and metrics live outside the repo, preventing agent tampering.
- Fully auditable results — Each accepted change is a separate commit and every experiment is logged in results.tsv.
- Stop anytime — You can halt the run without losing any already-accepted changes.
- Multi-language support — Works with Go, Python, TypeScript, JavaScript, Java, C#, Rust and Swift.
- Agent skill packages — Provides SKILL.md packages for Claude Code, Cursor, Codex CLI and any agentskills.io-compatible agent.
- Benchmark-driven loop — Runs interleaved A/B benchmarks, discarding noise and rejecting regressions automatically.
- Open source MIT license — Free to use, modify and integrate into your own workflows.
Use cases
- Run an AI agent on a repository and get proven performance gains.
- Validate that a suggested code change does not introduce regressions.
- Generate a commit history of performance-focused optimizations.
- Integrate with existing AI coding agents like Claude Code or Cursor.
- Measure and compare optimization impact across multiple languages.
autor3search vs alternatives
 autor3search |  Cursor |  Augment Code |  OpenCode |  Kiro | |
|---|---|---|---|---|---|
| Best for | Verified AI-driven code optimizations | AI-assisted code editing | Context-rich AI coding assistance | Terminal-based AI coding | Spec-driven agentic IDE |
| Pricing | Open source | Subscription | Subscription | Free | Subscription |
| DevHunt upvotes | 1 | 0 | 0 | 0 | 0 |
| Launched | Oct 2026 | — | — | — | — |
- autor3search vs Cursor: Cursor is an AI-enhanced editor, not a benchmark-driven verification harness.
- autor3search vs Augment Code: Augment Code provides deep codebase context but does not automatically benchmark changes.
- autor3search vs OpenCode: OpenCode is an AI terminal agent without built-in performance verification loops.
- autor3search vs Kiro: Kiro focuses on spec-driven development, not on measuring runtime performance of AI suggestions.
autor3search FAQ
How does autor3search ensure the AI cannot cheat the benchmark?+
Tests are frozen and restored before each run, and the performance metric lives in a compiled binary outside the repo, inaccessible to the agent.
What happens to changes the AI proposes that fail the test suite?+
The harness discards them; the repository is reset to the previous commit and the failure is logged.
Can I stop the optimization loop mid-run?+
Yes, you can stop at any time; any changes already kept remain in the repository.
Is the tool compatible with multiple programming languages?+
Supported languages include Go, Python, TypeScript, JavaScript, Java, C#, Rust and Swift.
Do I need to modify my codebase to use autor3search?+
No, you only add the SKILL.md package and run the provided commands; the tool works on existing repos without code changes.
Summarized by DevHunt from autor3search.dev · Oct 25, 2026. Details may change; check the official site.
Around the web
autor3search/javascript: Autonomous AI-driven performance optimization for any JavaScript repositoryGitHub
autor3search - Rust - Docs.rsdocs.rs
autor3search-rust on crates.iocrates.io
autor3search-java on Maven Repositorymvnrepository.com
autor3search-go on Libraries.iolibraries.io
autor3search on piwheelspiwheels.org
