Who Owns Your Agent Traces?
Agent traces are the best training data that exists. So who owns them? Notes from an AI+ dinner on the learning loop, RL, and cloud economics.
TL;DR
- AI agent traces are records of every step an agent takes, including tools used, dead ends, and recoveries.
- Traces are considered the best training material for improving AI agents, forming a 'learning loop' or 'flywheel'.
- Ownership of agent traces equates to ownership of the learning loop and potentially the future of a company.
- The enterprise, agent company, infrastructure layer, and model lab all have defensible claims to trace ownership.
- A researcher noted that models trained via reinforcement learning learn from their own attempts, not necessarily directly from traces (supervised fine-tuning).
- Customizing models for individual companies can break cloud economics by shifting from pooled capacity to dedicated, non-multi-tenant resources.
- AI agents produce permanent, timestamped records of actions, unlike humans, creating accountability and leading to arguments over 'who gets to hold the receipts'.
- The ability to find rare but critical failure points (0.01%) in agent operations is a key reason for the importance of traces.