AIRun: A Production Framework for Failure-Driven MCP Tool Evolution

Building tools for AI agents is fundamentally different from building human-facing APIs. AI agents make non-obvious, systematic mistakes that are invisible to traditional API observability: they omit critical query constraints, misinterpret ambiguous schema descriptions, and fail silently on timeout. We present a five-phase bootstrap methodology for developing AI-consumable tools by systematically observing agent failure patterns in production and iteratively evolving tool descriptions, schemas, and infrastructure in response. Applied to a production Model Context Protocol (MCP) server at eBay serving AI-driven incident investigation workloads, the methodology improved agent query success rate from 60% to 95%, reduced partition filter omission from 75% to 5%, and reduced investigation time from 20-30 minutes (experienced-practitioner benchmark) to under 3 minutes in documented case studies (N=3; see Section 5). A key contribution is the Closed-Loop Concurrent Agent Development Cycle: one agent session uses the tool in production while a parallel agent session reads structured failure logs and implements fixes - an AI-assisted feedback loop that delivered 15 pull requests in 4 weeks at 32 total human hours. We describe the failure taxonomy, tool description anti-pattern format, correlation ID observability schema, and three-stage async architecture that together constitute a principled engineering methodology for production AI tooling.