SeedX

🇺🇸 English · 🇨🇳 简体中文 · 🇯🇵 日本語

Every idea can grow into a system.

SeedX is a multi-agent system that turns a work or learning question into a complete learning package: executable, evaluable, and transferable.

Give it a question from Claude Code, Hermes, OpenClaw, or a mobile workflow. SeedX plans the learning goal, splits the work, builds the artifacts, evaluates them, repairs weak parts, and leaves a visible trail of agent collaboration, prompts, data flow, and outputs.

Preview

_{A completed SeedX run: task progress, agent handoffs, event timeline, and generated learning artifacts stay observable in one workspace.}

Why It Exists

SeedX has two jobs:

1. For learners: convert a messy question into a structured path you can actually follow.
2. For agent builders: test whether model agents can complete long, multi-step work inside a harness without human intervention.

It is also a testbed for long-task runs with Chinese-model agents, including MiniMax M2.5/M2.7 experiments.

The current harness is designed around three constraints:

• No human-in-the-loop during a run.
• Agent collaboration, prompts, state, and artifacts must be visible.
• Every learning package must be executable, evaluable, and transferable.

Quick Start

Recommended flow:

1. Copy your question body to the clipboard.
2. In Claude Code, Hermes, OpenClaw, or another connected agent entry point, send:

+ask

The hook saves the question under input/questions/, starts the orchestrator, opens the visualizer, and writes the final learning package under output/{english-topic-yymmdd-HHMMSS}/.

You can also start with an explicit file:

+start input/questions/{question-file}.md

Legacy and direct triggers are still supported:

seedx <question>
seed <question>
sx <question>
qtm <question>

qtm remains available as the legacy Question-to-Mastery trigger.

What You Get

Each run creates a project folder under output/:

output/{project}/
├── README.md
├── deliverables/
│   ├── question-brief.md
│   ├── domain-map.md
│   ├── learning-path.md
│   ├── exercises.md
│   ├── checkpoints.md
│   ├── application-plan.md
│   └── transfer-plan.md
├── _agent/
│   ├── learning-plan.md
│   ├── learning-contract.md
│   ├── learning-design-guide.md
│   ├── project-lessons.md
│   └── review-reports/
└── _run/
    ├── run-log.md
    ├── events.jsonl
    └── state.json

The learner-facing files are in deliverables/. Agent notes, evaluation reports, and runtime state stay in _agent/ and _run/.

How It Works

flowchart LR
    A[Question file] --> P[question-planner]
    P --> C[learning contract + design guide]
    C --> B[mastery-builder]
    B --> D[deliverables]
    D --> E[learning-evaluator]
    E -->|PASS| N[next task]
    E -->|FAIL| R[resume same builder]
    R --> E
    E --> V[events + state visualizer]

Loading

SeedX runs three fixed task units:

Task	Purpose	Outputs
task01	Frame the question	deliverables/question-brief.md, deliverables/domain-map.md
task02	Build the mastery path	deliverables/learning-path.md, deliverables/exercises.md, deliverables/checkpoints.md
task03	Apply and transfer	deliverables/application-plan.md, deliverables/transfer-plan.md

Each task is built, evaluated, and repaired for up to two rounds before the system moves on.

Visualization

The visualizer reads only runtime state, not learning artifact bodies:

./tools/open-visualizer.sh {project}

Without a project name, it opens the newest project under output/.

It is useful for inspecting the long-task run: which agent is active, what prompt was handed off, which files were produced, and whether each task passed.

Advanced Usage

Use +ask when you want the main orchestrator to receive only a file path. Use direct triggers like seedx <question> or qtm <question> when you are fine with the question body appearing in the original chat message.

For sensitive questions, use clipboard mode:

+ask

For manually controlled runs, create a question file under input/questions/, then start with:

+start input/questions/{question-file}.md

For Maintainers

• Orchestration protocol: AGENTS.md
• Claude Code protocol mirror: CLAUDE.md
• Output layout: docs/specs/output-artifact-layout.md
• Event protocol: docs/specs/harness-observability-events.md
• Run log format: docs/specs/run-log-format.md
• Repository hygiene: docs/specs/repository-hygiene.md
• Architecture decision: docs/adr/0001-question-to-mastery-architecture.md
• SeedX rename note: docs/release-notes/seedx-rename.md

Contact: X @CaoYuhaoCarl · Telegram @caoyuhaocarl · WeChat caoyuhaocarl