agentgradelisted

# agentgrade — production-readiness auditor for AI agents Grade an AI agent codebase against five pillars, cite evidence for every score, and hand back a short ranked list of what to build next. The output is a gap report, not a lecture: scorecard → verdict → top 3 gaps → offer to scaffold. ## Process ### 1. Scope the agent Identify what you're auditing before judging it: - Find the agent surface: LLM SDK imports (`anthropic`, `openai`, `google.generativeai`, `litellm`, `ollama`, Vercel `ai`/`@ai-sdk`), frameworks (`langchain`, `langgraph`, `crewai`, `autogen`, `pydantic_ai`, `llamaindex`, `haystack`, `dspy`, `semantic-kernel`, `smolagents`, `claude-agent-sdk`, `openai-agents`, Google ADK, Strands, `agno`, `letta`, MCP servers), entry points, tool definitions. - Count agents: single agent vs multi-agent (affects the orchestration pillar). - Detect RAG: vector stores, embedding calls, ingestion scripts (affects data foundation). - Establish deployment context — it calibrates the bar. Infer from the repo (Dockerfile + auth + customer-facing API ⇒ production intent; a CLI demo ⇒ demo). If genuinely ambiguous, ask once: internal tool, pilot with real users, or customer-facing production? ### 2. Read the rubric Read [references/rubric.md](references/rubric.md). It defines, per pillar: what to look for (grep/glob signal patterns), anchored score levels 0–5, and when a pillar is N/A. Scores come from the anchors, never from vibes. ### 3. Collect evidence Start