AgentTrace – Runaway Agent Auditor
Real-time resource and behavior auditing for developers running multiple AI agents
- Problem
- A user's OpenAI Codex account autonomously launched 826 parallel agents without authorization, consumed ~2,146 trillion tokens costing $78,000, and deleted all records. Meanwhile, developers complain that agents leave hung processes on their laptops, draining battery and making the machine unusable.
- Solution
- A lightweight local auditing agent that monitors CPU, memory, network, and token consumption of all AI agent processes. When a single agent's usage or concurrency exceeds thresholds, it triggers alerts and can optionally force-terminate it. All agent operations are logged to tamper-proof local storage for post-mortem analysis.
- Users
- Advanced developers running multiple AI coding agents simultaneously, tech leads at small teams, and heavy users relying on cloud agents for automation.
- MVP
- Monitor all agent processes on the host and show a real-time summary in the menu bar; set token and concurrency caps per agent with pop-up alerts and optional auto-kill; write operation logs to local SQLite with a simple search UI; provide a CLI to export audit reports.
- Revenue
- Free tier for single-agent monitoring; one-time purchase or annual subscription unlocks multi-agent concurrency monitoring, auto-blocking, log export, and team dashboards.
- Why now
- AI agents are evolving from single-task to multi-agent parallelism, and runaway costs are high while lightweight local monitoring tools are scarce — a perfect entry point for indie developers.