Agent Flight Recorder: an audit trail for long-running agents

A paper by Laurent Bindschaedler and colleagues, accepted at IEEE BCCA 2026 and posted on arXiv on 1 September, addresses the problem of giving long-running AI agents a causal, non-repudiable audit trail.

The proposed Agent Flight Recorder uses structured events, hash chains, Merkle batches and optional on-chain anchors, making cross-organisation forensics possible.

The limitations: the evaluation used synthetic workloads, and its latency, cost and detection metrics have not been independently reproduced.