LLM-based agents operate by making sequences of executable decisions, but the resulting trajectories carry little to no information about which specific agent or policy generated them. This makes it difficult to establish provenance, verify ownership, or detect unauthorized reuse of a model's behavior.
To address this, the paper introduces a concept called sequential behavioral watermarking. The idea is to embed identifiable markers into the agent's decision sequence itself, so that the originating model can be recognized from its behavior alone. The approach is described in a recent arXiv preprint.
Such watermarking could help developers protect their agents and provide a basis for auditing and accountability. The paper does not yet detail the full method or its limitations, but it highlights a growing concern as LLM agents become more autonomous and widely deployed.