AI agent impact analysis

Most teams running AI agents and MCP infrastructure have no single place to ask “what breaks if I change this?” The answer lives scattered across config files, agent definitions, scripts, and tribal knowledge. By the time someone needs the answer, usually right before a risky change, it has to be reconstructed by hand under time pressure.

Dependency relationships in AI infrastructure are rarely declared in one place. A credential might be referenced by name in one file and consumed through an SDK import in another; an agent might depend on a tool it discovers at runtime rather than importing directly. A relationship also isn't binary: some are certain (a direct import), some are structural signals (a cron job that plausibly triggers the agent), and some can't be resolved at all without guessing.

Wirecheck runs deterministic detection first, regex and parser-based extraction across supported file types, then uses an LLM only to classify genuinely ambiguous cases, never to invent a relationship the deterministic pass didn't find. Every result carries a confidence level (HIGH, MEDIUM, or UNKNOWN) and evidence pointing at the file and line it came from. See Supported detections for exactly what is and isn't covered today.

wirecheck impact customer-support-agent --operation MODIFY --json

{
  "component": { "name": "customer-support-agent", "type": "AGENT", "confidence": "HIGH" },
  "direct_dependants": 0,
  "indirect_dependants": 2,
  "orphaned_integrations": 1,
  "unknown_relationships": 0,
  "indirect": [
    { "name": "zendesk-mcp", "type": "MCP_SERVER", "relationship": "USES" },
    { "name": "slack-notify", "type": "WEBHOOK", "relationship": "TRIGGERS" }
  ]
}

Zero direct dependants, two indirect ones, and one orphaned integration worth a second look before modifying the agent.

Every relationship is traceable to a repository, file, and line, with a confidence level attached. “2 indirect dependants” is not a count pulled from a heuristic score; it is two specific locations you can open and check.

A bare LLM API call is deliberately not classified as an agent; Wirecheck requires stronger signals (an agent SDK or class, tool definitions, a system prompt plus model and tools, or an agent loop) before assigning HIGH confidence. Weaker combinations get MEDIUM confidence rather than being silently treated as certain, and anything the deterministic pipeline or classifier can't resolve stays UNKNOWN.