How to Build AEO Tracking for Coding Agents

As AI transforms how information is searched, companies are realizing the importance of AEO (AI Engine Optimization), understanding how LLMs search and summarize web content.

While traditional standard chat models can be analyzed by examining API calls through AI Gateway, coding agents that operate in terminals or IDEs require a completely different approach. Since coding agents need a full development environment including file systems, shell access, and package managers, securing execution isolation and observability is the key challenge.

To solve this, Vercel built a tracking system leveraging Vercel Sandbox. This system uses ephemeral Linux MicroVMs that spin up in seconds, following a 6-step lifecycle:

• Sandbox creation: Launch a MicroVM with a specific runtime (Node.js, Python, etc.)
• Agent CLI installation: Globally install an agent such as @anthropic-ai/claude-code
• Credential injection: Route all LLM calls through Vercel AI Gateway for unified logging and cost tracking
• Agent execution: Run each CLI's unique commands as shell commands
• Transcript capture: Extract records of tools used, whether web search occurred, recommendations made, etc.
• Resource cleanup: Terminate the sandbox and release resources

Thanks to this uniform lifecycle, adding a new agent can be instantly integrated into the system by simply adding a configuration object, without complex implementation.