AgentSkillsCN

agent-specification-standard

在构建或审查 LangGraph 智能体、推理引擎，或集成 MCP 的认知工作流时，可应用 ASS-001。

SKILL.md
--- frontmatter
name: agent-specification-standard
description: Apply ASS-001 when building or reviewing LangGraph agents, reasoners, or MCP-integrated cognitive workflows.

Agent Specification Standard (ASS-001)

Use this skill when implementing or auditing probabilistic "Reasoner" agents that use LangGraph, MCP tools, and checkpointer-based state.

When to Use

  • •Designing or implementing LangGraph-based agents
  • •Integrating agents with MCP (Model Context Protocol) for tool discovery
  • •Defining agent state schemas, system prompts, or safety guardrails
  • •Setting up agent observability (OTel, reasoning logs) or HITL interrupts

Instructions

  1. •State & determinism: Define a strictly typed State Schema (Zod). Use a checkpointer (Postgres or Temporal-backed) for resumable agentic loops. Keep nodes as pure as possible; offload side effects to OCS Capabilities.
  2. •Tools: Do not hardcode API clients. Bind tools via MCP; discover Blueprints and Capabilities from the MCP Registry at runtime.
  3. •Safety: Version system prompts as code; include a "Negative Constraint" block. Define max_iterations and token_limit per cycle. Trigger HITL interrupts for tools marked RESTRICTED in OCS.
  4. •Observability: Nest LangGraph spans under the parent OTel Trace ID. Capture "thought process" in structured logs and strip from user-facing responses.

For the full normative standard and TypeScript/LangGraph contract, see references/agent-specification-standard.md.