Overview

Synthetic Scientists is a plugin-driven research runtime for coding agents. The host agent turns an objective into acceptance criteria and an evaluator, creates a shared work DAG, assigns personas, launches parallel builders, and returns a verified commit with a complete evidence record.

Session

1. scientist launch -c task.yaml creates the session, worktrees, shared commons, work DAG, and evaluator daemon.
2. Workers claim DAG nodes and implement isolated candidates.
3. scientist eval --node <id> -m "..." commits and evaluates the candidate.
4. Scores, decisions, evidence, and follow-up work attach to the DAG.
5. Coordinators assign additional work when more than eight workers are active.
6. The host orchestrator verifies and promotes the selected commit.

Navigation

Components

• Host orchestrator clarifies intent, designs evaluation, seeds the work DAG and personas, operates the session, and verifies delivery.
• Workers are stock coding-agent CLIs running in isolated git worktrees.
• Coordinators manage one group of four workers above the configured threshold.
• Evaluator daemon checks every submitted commit at its exact state.
• Shared commons contains experiments, evidence, methods, checkpoints, and the global coordination DAG.
• Labs maintain separate search populations with periodic rotation.
• Cadence injects reflection, synthesis, redirect, and curation prompts.
• Live interface shows session health, rankings, work, evidence, and agent activity.

Core commands

scientist new my-task
scientist check my-task
scientist launch -c my-task/task.yaml

scientist overview
scientist results --recent
scientist work list
scientist dashboard

scientist promote <hash> --branch verified-result