Universal Agentic Harness Protocol & Multimodal Orchestration Standard (v2.2)
MetaScout is a Host-Agnostic Universal Agentic Harness Protocol (v2.2). It serves as a Superset of traditional Agent Package Managers (such as Vercel Labs' find-skills).
While traditional tools only search and install static skill packages, MetaScout provides a complete autonomous problem-solving loop: Pre-Plan Confirmation ➔ Search-Driven Autonomous Research & Quality Guard ➔ Human-in-the-Loop Pre-flight Review ➔ Hot-Swappable Execution & Adaptive Multimodal Presentation ➔ Experience Distillation.
| Feature / Dimension | find-skills (Vercel Labs) |
MetaScout Protocol (v2.0) |
|---|---|---|
| Open Ecosystem Index | skills.sh CLI (npx skills find/add) |
✅ Integrated (skills.sh, Smithery, GitHub, Web Search) |
| Quality & Security Guard | Manual Stars / Audits check | ✅ Automated Quality & Security Filtering (Installs ≥ 1K, Verified Authors) |
| Methodology Synthesis | ❌ Limited to existing static skill packages | ✅ Search-Driven Autonomous Methodology Synthesis |
| Human-in-the-Loop Review | ❌ Direct installation without pre-flight plan | ✅ Pre-flight Blueprint Gatekeeper Review |
| Environment Hygiene | ❌ Permanent package pollution | ✅ Sandboxed Hot-Mounting & Self-Destruct |
| Adaptive Output | Monochromatic Text / Markdown | ✅ Multimodal Rich Media (Direct Purchase Links, Map Nodes, Storyboards, HTML/PPT) |
MetaScout addresses the critical dilemma in modern LLM Agent design: Context Bloat vs. Handoff Loss. When an agent tries to manage too many tools, its reasoning degrades. If we use permanent subagents, implicit context is lost during handoffs.
MetaScout solves this by adopting a Three-Tier Constitutional Architecture:
- Strategic Layer (Main Agent as Route Generator): The main agent NEVER executes tool installation or dirty work. It stays pristine, acts as a "Meta Brain", parses the goal, and synthesizes a Pre-flight Blueprint (DAG).
- Tactical Layer (JIT Registry & Gatekeeper): Instead of loading tools into prompt, tools are searched Just-In-Time from registries (
skills.sh/Smithery). The user acts as the Gatekeeper to approve the Blueprint. - Execution Layer (Disposable Sandbox Subagents): Executing the blueprint is offloaded to a temporary subagent that runs the tools and performs Visual QA/Self-Verification. Once done, it passes back clean, final artifacts and self-destructs, eliminating handoff loss while protecting the main context window.
┌───────────────────────────┐
│ 0. Pre-Research Plan & │
│ 1. Capability Discovery │
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ 2. Explore → Evaluate → │
│ Execute Blueprint │
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ 3. Human-in-the-Loop │
│ Pre-flight Gatekeeper │
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ 4. Hot-Swappable Sandbox │
│ Execution & Visual QA │
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ 5. Experience Distillation│
│ & Subskill Saving │
└───────────────────────────┘
- Pre-Plan & Discovery: Shows research strategy before searching. Checks memory for past subskills and audits capability gaps.
- Explore & Evaluate: Prioritized JIT search (PATH -> skills.sh -> Smithery -> GitHub). Synthesizes a Blueprint comparing tool candidates.
- Gatekeeper Review: Presents the Blueprint to the user for explicit authorization.
- Execution & QA: Hot-mounts tools in a sandbox. Enforces Visual QA (e.g. converting PPT to images to check for text overflow) before adaptive multimodal delivery (file:// paths).
- Experience Distillation: After successful execution, the agent self-scores. High-scoring DAGs and methodology notes are saved as lightweight "Subskill Blueprints" in
memories/for future reuse, preventing token waste on repeated tool-chaining.
MetaScout is fundamentally a Prompt-based Protocol, which means it is host-agnostic and can be installed into any mainstream AI Agent without requiring complex plugins.
You can inject the MetaScout Harness into tools like Cursor, Claude Code, ChatGPT, Grok, or Aider instantly:
- Open your agent's custom instructions, system prompt, or rule file (e.g.,
.cursorrules,claude.md, or Custom GPT instructions). - Copy the entire contents of
skills/metascout/SKILL.mdand paste it into the file. - Your agent is now upgraded with the MetaScout v2.2 Harness Protocol and will strictly follow the Five Pillars for any autonomous task.
If you are using an environment that supports Antigravity plugins:
git clone https://github.com/Simple53/MetaScout.git ~/.gemini/config/plugins/MetaScout(Refer to ADAPTERS_GUIDE.md for detailed platform-specific adapters).
- HARNESS_SPEC.md - Universal Protocol Specification (v2.0 Superset)
- ADAPTERS_GUIDE.md - Multi-Agent Framework Integration Guide
- zh/README.md - 中文说明文档
Distributed under the MIT License. See LICENSE for details.