Synthetic Scientists

Scale inference-time compute for open-ended research.

Turn a research objective and evaluator into long-running coding agents that run experiments, accumulate evidence, and return a verified result.

Start · Research loop · Install · What is open · Docs

---

Start a research session

Install the CLI and the plugin for your coding agent, then describe the research outcome:

Use Synthetic Scientists to improve this solver.
Preserve correctness and minimize the benchmark runtime.

The host agent reads the repository before asking anything. It proposes one compact research contract:

Objective      Reduce benchmark runtime
Metric         paired runtime ratio, maximize
Correctness    existing tests + held-out cases
Isolation      srt; network disabled
Compute        4 autonomous workers · no automatic stop
Deliverable    verified branch + reproduction command

Launch, or change any line?

After approval, workers run the research loop autonomously. The host monitors health and budget, verifies the selected commit against a matched baseline, and imports the branch into the source repository.

The research loop

objective + evaluator + compute budget
                  │
                  ▼
        registered baseline
                  │
     ┌────────────┼────────────┐
     ▼            ▼            ▼
 retrieve      propose     local checks
     ▲            │            │
     └──── shared evidence ◄─ evaluate
                  │
                  ▼
     repeated candidate + baseline
                  │
                  ▼
           verified branch

Workers control retrieval, hypotheses, local testing, and evaluation timing. Synthetic Scientists supplies isolated worktrees, scored commits, persistent research memory, user-directed compute, health management, and final verification.

Why scale inference-time compute?

• Search breadth: independent workers explore different mechanisms and starting points.
• Search depth: long-lived sessions iterate through implementation failures instead of stopping after one generation.
• Knowledge accumulation: experiments, failed approaches, notebook evidence, and reusable methods remain available to every worker.
• Verification: search-time leaders are repeated against the registered baseline before delivery.

Representative uses include algorithm discovery, systems optimization, machine-learning experimentation, and scientific software.

Install

CLI

npm install -g synthetic-scientists
scientist login

The launcher provisions the core into ~/.synthetic-scientists/; project dependencies remain untouched.

Claude Code

/plugin marketplace add synthetic-sciences/scientist
/plugin install scientist@scientist-marketplace

Codex

codex plugin marketplace add synthetic-sciences/scientist
codex plugin add scientist@scientist-marketplace

Cursor

agent plugin marketplace add \
  https://github.com/synthetic-sciences/scientist.git \
  --git-ref main

Run /plugin, select scientist, and choose User or Project scope.

OpenCode and shared skills

git clone https://github.com/synthetic-sciences/scientist \
  "$HOME/.local/share/synthetic-scientist"
"$HOME/.local/share/synthetic-scientist/plugin/install-skills.sh"

The durable .agents/skills installation is also discovered by Cursor and Codex.

Scale presets

Preset	Workers	Organization
solo	1	One autonomous trajectory
team	4	Autonomous peers + shared evidence
diverse	8	Peers + adaptive search biases
hierarchical	16+	Coordinator groups

Presets do not add automatic stops. Score, evaluation, and wall-time conditions remain available when the user explicitly requests a bounded campaign.

Verification and delivery

scientist verify <hash> --repetitions 3
scientist promote <hash> \
  --branch research/verified-result \
  --into /path/to/source-repository

verify regrades the clean candidate and matched run baseline under the registered evaluator. promote --into transfers the verified branch without checking it out or modifying the user's working tree.

What is open

plugin/
├── skills/
│   ├── scientist-orchestrator/  intent → contract → research → verification
│   ├── scientist-bootstrap/     installation and source snapshots
│   ├── scientist-evaluator/     evaluator and acceptance-contract authoring
│   ├── scientist-operator/      session operation and recovery
│   └── scientist-profiles/      runtime profiles
├── agents/                      task authoring and session diagnosis
├── hooks/                       non-provisioning session context
└── host manifests/              Claude Code, Codex, and Cursor

launcher/                        npm launcher source
install.sh                       no-Node installation path

The skills, launcher, and installer are Apache-2.0. The separately distributed core requires access.

Direct CLI

scientist new my-task
scientist check my-task
scientist launch -c my-task/task.yaml agents.count=4

scientist overview
scientist results --recent
scientist inspect <hash> --diff
scientist verify <hash>
scientist promote <hash> -b research/result --into /path/to/source

Full reference: documentation.