Agent Health Check

Governance scoring for AI agent system prompts, 0 to 100

agent_health_check scores any AI agent's system prompt on 17 governance best practices, returning a score from 0 to 100 with specific findings. It detects authority leaks, silent inference, missing audit trails, and other anti-patterns that make agents unsafe. Free to use, no API key required.

Pass your agent's full system prompt text. You get back a governance score, a list of specific issues found with severity ratings (CRITICAL, WARNING, INFO), and actionable recommendations to fix each one.

What It Checks

agent_health_check scans for 17 governance anti-patterns across six categories.

Authority and Scope

• Authority leaks -- the agent claims or assumes authority it was not explicitly given
• Scope creep -- instructions that expand beyond the agent's stated purpose
• Vague delegation -- unclear boundaries on what the agent can and cannot do
• Unconstrained tool use -- no limits on which tools the agent may invoke or when

Transparency

• Silent inference -- the agent makes decisions without explicit reasoning or user awareness
• Missing identity disclosure -- no indication the user is interacting with an AI
• Hidden reasoning -- conclusions presented without showing how they were reached

Accountability

• Missing audit trails -- no mechanism for logging or reviewing agent decisions
• No human escalation path -- no defined process for handing off to a human
• Missing error handling -- no guidance for what the agent should do when things go wrong

Security

• Prompt injection exposure -- system prompt lacks defenses against injection attacks
• Data boundary violations -- no restrictions on what data the agent can access or share

And More

The full check covers 17 anti-patterns including missing output constraints, feedback loop risks, unversioned prompt changes, and others. Each finding includes a severity level and a specific fix.

How to Use It

Connect to the Agentic Platform MCP server. Then call agent_health_check with the full text of your agent's system prompt.

Step 1: Connect

claude mcp add agentic-platform --transport http https://api.asalvocreative.com/mcp

{
  "mcpServers": {
    "agentic-platform": {
      "url": "https://api.asalvocreative.com/mcp"
    }
  }
}

npx -y @smithery/cli@latest install agentic-platform--andysalvo.run.tools --client claude

Step 2: Call the Tool

agent_health_check({
  system_prompt: "You are a helpful assistant. Do whatever the user asks. Use any tools available to complete tasks efficiently."
})

Example Output

Agent Health Check Report
=========================
Score: 28/100

CRITICAL  No authority boundaries defined
CRITICAL  "Do whatever" grants unlimited scope
WARNING   No error handling guidance
WARNING   No audit or logging mentioned
WARNING   No human escalation path
INFO      No identity disclosure

Recommendations:
  1. Define explicit scope boundaries
  2. Add authority transfer protocols
  3. Include error handling instructions
  4. Add decision logging requirements
  5. Define human escalation triggers
  6. Disclose AI identity to users

A score of 28 means critical governance gaps. Most production agent system prompts score between 40 and 70 on first check. Following the recommendations typically raises the score above 80.

Interpreting Your Score

80 to 100: Strong governance. The system prompt defines clear boundaries, includes audit mechanisms, and handles edge cases.

50 to 79: Partial governance. Some safeguards exist but gaps remain. Review the specific findings to close them.

Below 50: Critical issues. The agent likely has authority leaks, no audit trail, or missing error handling. Address CRITICAL findings first.

About the Agentic Platform

The Agentic Platform is an MCP server providing free agent diagnostics. In addition to the health check, it offers an MCP linter, cost estimator, and service trust evaluator -- all free, no API key required.