How we built AEO tracking for coding agents

This post explains how Vercel built an AI Engine Optimization (AEO) tracking system specifically for coding agents like Claude Code and OpenAI Codex.

• •Coding agents behave differently from standard chat models, performing web searches in ~20% of prompts and requiring a full dev environment (filesystem, shell, package managers)
• •Vercel Sandbox provides ephemeral Linux MicroVMs that follow a 6-step lifecycle: create sandbox, install agent CLI, inject credentials, run agent, capture transcript, tear down
• •AI Gateway is used to centralize LLM call routing by overriding provider base URLs (e.g., ANTHROPIC_BASE_URL) via env vars, enabling unified logging and cost tracking
• •Each agent produces transcripts in different formats (Claude Code writes JSONL to disk, Codex streams JSON to stdout), so a 4-stage normalization layer was built to unify them
• •Tool name normalization maps agent-specific names (Read/read_file/read) to ~10 canonical names for consistent analysis across agents