Synthetic Scientists: Intent‑to‑Evidence
Software Research
System Note · August 2026
Abstract
Synthetic Scientists turns a measurable software objective into a verified research session. A host coding agent clarifies intent, defines acceptance criteria, builds the evaluator, creates a shared work DAG, and assigns task-aware personas. Parallel workers test isolated changes in git worktrees; an evaluator daemon scores every commit at its exact state. Results, decisions, evidence, and follow-up work remain attached to the DAG. Above eight workers, one coordinator is added per group of four. The selected result is independently checked and promoted as a normal branch with the session record required to reproduce it.
1Research session
A session contains a host orchestrator, isolated coding-agent workers, optional coordinators, and an evaluator daemon. The host owns intent and delivery. Workers own scoped DAG tasks. Coordinators manage groups of four workers when the session grows. The evaluator applies correctness and scoring rules from a hidden, reproducible environment.
Shared files under .scientist/public/ form the session commons: scored experiments, notebook entries, reusable methods, checkpoints, persona assignments, and the work DAG. Labs can maintain separate populations while every lab reads the same coordination state.
2Install
The npm package synthetic-scientists installs the scientist command; scientist login stores your license key and provisions the core into an isolated, managed environment.
npm install -g synthetic-scientists
scientist login <YOUR-LICENSE-KEY>Requirements: Node 18+, git, and one supported coding-agent CLI. Python is managed automatically.
3Configure
Detect local runtimes, create named profiles, and verify that each profile can authenticate and answer:
scientist configure
scientist profiles verify4Launch
Scaffold a task, check its evaluator, and launch a session:
scientist new my-task
scientist check my-task
scientist launch -c my-task/task.yamlTable 1: Operating a live session.
| Command | Purpose |
|---|---|
scientist overview | session health, rankings, and work state |
scientist results --recent | recent scored experiments |
scientist inspect <hash> --diff | one experiment and its code change |
scientist work list | the shared task-and-commit DAG |
scientist dashboard | the live interface |
Evaluation design is documented in Evaluation Design. Every selected result can be promoted with scientist promote.
5Plugin
Install the plugin in a host coding agent and state the objective directly. The scientist-orchestrator skill handles evaluation design, coordination, verification, and delivery [3].
/plugin marketplace add synthetic-sciences/scientist
/plugin install scientist@scientist-marketplaceReferences
- [1]Quickstart: install, configure, check, and launch. docs/quickstart
- [2]System: worktrees, evaluation, commons, DAGs, personas, and labs. docs/system
- [3]Plugin orchestration from Claude Code, Codex, or OpenCode. docs/playbooks
- [4]Source code. github.com/synthetic-sciences/scientist