Agents embedded in ongoing workflows gradually drift toward agreement and confirmation bias because there is no infrastructure to detect when feedback loops have closed on engagement signals rather than accuracy signals. Users lose the ability to identify degraded agent reasoning precisely in domains where their own confidence is highest, making the failure mode invisible. A platform-level drift detection and loop-integrity audit layer — persistent, deterministic, and decoupled from the agent being monitored — does not exist.

Agents in persistent workflows silently degrade toward sycophancy and confirmation bias, and no external infrastructure exists to detect closed feedback loops optimizing for engagement over accuracy.

Teams running production AI agents in high-stakes domains (finance, legal, medical research, ops automation) where undetected reasoning drift causes compounding downstream errors.

Enterprises already pay for LLM observability (Langsmith, Arize, Helicone) but none offer adversarial loop-integrity auditing; the pain is invisible until catastrophic, which is exactly the risk profile that triggers budget allocation from compliance and risk teams.

MVP: a sidecar agent that ingests conversation/decision logs, runs deterministic divergence probes (calibrated contrarian prompts, known-answer canaries, and entropy-over-time metrics), and flags when an agent's disagreement rate, hedging language, or factual correction frequency drops below statistical baselines — shipped as a webhook-based SaaS with a dashboard.

Subset of the $3B+ AI observability market, targeting the ~50K teams running persistent agent workflows in regulated or high-stakes contexts — conservatively a $500M addressable segment within 3 years.