Persistent AI agents in research: 96-day study reveals cache-dominant workflow

Anas H. Alzahrani's implementation case study (arXiv:2605.26870) is the first systematic look at what happens when large language model agents operate persistently in an academic research setting—not as one-off queries but as durable assistants with memory, tools, scheduled routines, and safety protocols. Over 96 days (January–May 2026), the single-investigator environment ran 17 specialized agent roles, produced 75,671 telemetry records, and logged 579.7 active hours. The system consumed 73.95 million tokens, of which a staggering 82.9% were cache reads—meaning the vast majority of reasoning reused prior context rather than re-computing from scratch.

The paper introduces PARE-M (Persistent Agentic Research Environment Measurement), a framework covering six dimensions: architecture, utilization, artifact production, resource use, reproducibility, and governance. Key findings include 482 output-proxy events and 889 failure/verification/correction events, alongside 502 memory files and 57 skill files. The cache-dominant pattern has a profound implication: future AI pricing may shift from cost-per-token to cost-per-completed-artifact, since cached reads are nearly free. This changes how researchers budget AI compute and opens the door to long-running, self-improving research agents that learn and adapt without exploding costs.