TELEOS: VALUE-AWARE ADAPTIVE TELEMETRY FOR PRIVACY-CONSTRAINED AND CAUSALLY USEFUL OBSERVABILITY IN AGENTIC AI SYSTEMS

Agentic artificial intelligence systems generate model, prompt, retrieval, memory, tool, policy, security, application, and infrastructure telemetry, yet current observability practice largely treats collection as a static configuration problem. This study introduces TELEOS, a decision-theoretic framework that treats telemetry acquisition as an online resource-allocation decision. TELEOS defines Telemetry Value of Information (TVoI) from expected entropy reduction, expected reduction in diagnostic decision loss, provenance/reconstruction value, operational consequence, and penalties for collection cost, privacy exposure, latency, security sensitivity, and retention burden. A four-level escalation policy selects fidelity under hard governance constraints and explicitly excludes inaccessible proprietary chain-of-thought. A provenance graph links user intent, planner decisions, model calls, retrieval, memory, authorization, tools, infrastructure, verifiers, and outcomes. To test the decision rule without fabricating production evidence, we implemented a controlled synthetic fault-diagnosis experiment with 14 latent states, 16 telemetry channels, 20 random seeds, and 800 episodes per seed for each of five policies. TELEOS achieved 72.27% fault-localization accuracy (95% CI 71.49–73.05%) at mean normalized telemetry cost 0.492, versus 81.04% at cost 1.910 for static-maximal collection. Against threshold adaptation, TELEOS improved accuracy by 8.16 percentage points while reducing cost by 8.38%. The results establish proof-of-concept for value-aware telemetry selection, while real-agent validation remains future work.