Open Discovery

Auto-optimize your code

The auto-optimize skill helps a coding agent optimize any system, algorithm, or codebase to run faster while preserving correctness and measuring the real hardware/workload tradeoffs.

Copy-paste prompt

Use this prompt to start an optimization session:

Install and use the `auto-optimize` skill from this repository:

https://github.com/vukrosic/open-discovery/blob/main/.agents/skills/auto-optimize/SKILL.md

Read it and start.

---

Open Discovery is a free, open-source collection of research and optimization workflows for coding agents. It turns one human idea, question, paper, feature, or system to improve into evidence-producing work.

Our immediate goal is to collect and design useful, fully digital experiments across science and engineering. We are building the library of questions, protocols, measurements, and evidence standards first. The next stage is an autonomous harness that can run these experiments, solve them, or search for improvements without losing track of what it actually proved.

The clearest product today is simple: open Open Discovery in Codex, Claude Code, or another capable coding agent; provide code plus a benchmark; and let the agent search for verified improvements. The repository and workflows are free. We will consider hosted execution only after real usage shows which operational problems people repeatedly need solved.

Install

git clone https://github.com/vukrosic/open-discovery.git
cd open-discovery

Open the cloned folder in Codex or start Claude Code from that folder. No Open Discovery server or package installation is required; individual workflows may use project-local dependencies when their task needs them.

Choose a workflow

Open this repository with a file-capable coding agent and choose a workflow:

• Discover: $discovery-engine, $find-ai-research-direction, $literature-review, $deep-strategy-research
• Design and analyze: $design-scientific-study, $curate-research-dataset, $analyze-scientific-data, $formalize-math
• Build and optimize: $paper-implementer, $build-benchmark, $evolve-program, $optimize-gpu-kernel, $optimize-inference, $optimize-agent
• Test and verify: $feature-tester, $red-team-agent, $audit-research-result
• Operate and extend: $package-research, $research-cockpit, $find-startup-skill-ideas, $build-skill

See all workflows and examples, or simply describe what you want in ordinary language. For example:

Optimize this code without changing its outputs. pytest must pass and python benchmark.py should become faster.

Examples

Current autonomous experiments

These small examples show the kinds of computer-only experiments Open Discovery is collecting and designing:

• Algorithm optimization: try a new version of an algorithm and measure whether it is faster while still giving the right answer.
• Computational biology: test a data-analysis method on gene-like data and check whether its predictions improve.
• 2D Ising physics simulation: simulate a simple magnetic system and compare two computer methods for sampling it.
• Public-assay hit triage: rank compounds from public EGFR assay data for a hypothetical follow-up screen and measure top-list enrichment.
• PID robotics control: tune a simulated robot controller and compare accuracy and stability on held-out disturbances.

Each example keeps the question, protocol, code, measurements, and generated results together. These are digital demonstrations, not claims about real materials, patients, or production systems.

Qwen3 Prompt Lookup Robustness is a research repository produced by AI agents using Open Discovery. The system autonomously investigated how to run Qwen3-0.6B faster on an Apple-Silicon MacBook, executed reproducible benchmarks, and found that fixed two-token prompt lookup was 30.4% faster than ordinary greedy decoding while producing exactly the same tokens.

Open Discovery records the request once and handles routine research decisions without making the researcher manage forms, files, agents, or approvals.

Architecture

Human brief
└── Initiative
    ├── Project A
    │   └── reviews, code, experiments, evidence, verification
    ├── Project B
    ├── Project C ...
    └── One canonical GitHub-ready repository artifact

An initiative is everything generated from one human request. A project is one independently testable question or engineering approach. Initiative leaders generate and compare projects; explorer agents own individual projects. A separate Scientific Reviewer can judge mature results using field-appropriate criteria, and a Research Communicator can turn accepted claims into accurate public drafts without publishing them.

Each initiative produces one canonical GitHub-ready repository artifact for the initiative as a whole, not one repository per project. A Repository Artifact Builder assembles inspected project outputs into that single package and keeps updating it as the initiative develops. It may include runnable setup, representative positive and negative results, prompts, provenance, and guidance for another human or agent to reproduce and continue the work. Local packaging does not automatically create or publish a GitHub remote.

Live work is organized as:

initiatives/<initiative>/
├── BRIEF.md
└── projects/
    └── <project>/

BRIEF.md is the only required research filename. Agents choose the remaining files, tools, code, and project organization according to the work. Open Discovery deliberately has no blank research templates and no fixed runtime.

What the harness provides

• AGENTS.md — the harness's operating behavior
• WORKFLOWS.md — user-facing workflow catalog and examples
• agents/ — research hierarchy, review, communication, and repository-artifact prompts
• .agents/skills/ — autonomous research and focused skills
• skill-lab/ — experimental skills loaded explicitly by path
• docs/ — evidence, collaboration, and operating guidance
• docs/AUTO-LAB.md — full computer-cycle auto-lab rule
• docs/AUTO-LAB-TEST-CASES.md — digital closed-loop dogfood case specs (runs stay out of git)
• docs/EXPERIMENT-CATALOG.md — experiment families and example tasks users and companies can set
• docs/EXPERIMENT-CONTRACT.md — the seven questions every experiment must answer
• research-modes/ — optional field-specific guidance

The harness provides prompts, responsibility boundaries, evidence standards, and durable-state principles. Agents generate the actual research structure and code they need.

For a continuing multi-initiative lab, local ignored files such as lab/MISSION.md and lab/CONSTRAINTS.md can define that lab's purpose and resource limits. They configure one lab instance and do not change Open Discovery's general behavior for other researchers. This workspace lab binds the auto-lab computer-only rule there; live initiatives under initiatives/ remain gitignored.

Default autonomy

Auto mode is the default for local, zero-cost, non-destructive research implied by the user's request. Spending, outside compute, publication, external communication, private access, credentials, and destructive actions still require explicit authority.

Current modes

Open Discovery includes guidance for AI and machine learning, mathematics, biology, cross-domain algorithm optimization, and evaluator-driven program evolution. The architecture also supports engineering and other evidence-driven work without pretending specialized guidance exists where it does not.

Limits

Open Discovery is a prompt-driven harness, not a server or cloud scheduler. A completed paper or result is not automatically correct, peer reviewed, novel, or published; claims remain limited by their preserved evidence.

License

MIT