Costile is an AI agent diagnostic proxy for teams building production AI workflows. As companies start using AI agents for support, research, coding, internal operations, and automation, they run into a new kind of infrastructure problem: the agent may technically work, but nobody knows when it is wasting money, looping, retrying badly, using the wrong model, or silently drifting outside expected cost behavior. Most teams can see their total AI bill after the fact. That is too late. What they need is operational visibility into which agent caused the spend, what pattern caused it, how severe it is, and what action should be taken. Costile sits between a company’s AI agents and model providers such as Anthropic. It routes requests through a proxy, logs only metadata such as model, tokens, cost, timestamps, agent ID, session ID, and stop reason, and turns that metadata into diagnostics. Prompts and responses are not stored. The product shows teams a command-center view of their AI agents: which agents are healthy, which are critical, what they spent in the selected time window, and what issue is driving the risk. When an agent is flagged, Costile gives a diagnostic report with the incident window, request log, baseline comparison, projected monthly exposure, and recommended fix. The problem we are solving is not just “AI is expensive.” The deeper problem is that AI agents are becoming autonomous software systems, but companies are still monitoring them like simple API calls. A bad agent loop, retry cascade, model-routing mistake, or runaway request can create cost, reliability, and governance risk before anyone notices. Costile helps engineering, product, and operations teams answer three questions quickly: What is going wrong? Which agent or session caused it? What should we do now? The initial product focuses on AI cost diagnostics, haywire detection, alerts, audit logs, API key management, and enterprise-ready reporting. Over time, Costile can become the observability and control layer for AI agent operations.