Current agent monitoring infrastructure detects capability gains but is structurally blind to gradual capability decay — including API changes, stale priors, data degradation, and approval drift. Agents continue operating with degraded performance and misaligned confidence because no tooling exists to continuously calibrate internal state against external ground truth. This asymmetry means failures appear sudden only because degradation was invisible, creating systemic reliability risk at scale.

Agent monitoring tools catch crashes and errors but miss slow capability rot — stale data, drifting API schemas, degrading model accuracy — so failures look sudden when they were actually gradual and preventable.

Platform engineering and MLOps teams running 10+ production AI agents at companies like fintech firms, e-commerce platforms, and SaaS companies where silent agent degradation has direct revenue impact.

Teams already pay $50K-500K/yr for observability (Datadog, Arize, LangSmith) but still get blindsided by slow-burn agent failures; Decay Radar fills a structural gap these tools weren't designed for, turning invisible drift into scored, actionable alerts before incidents happen.

MVP: a sidecar agent that periodically runs capability probes — synthetic ground-truth challenges, API schema diff checks, confidence calibration tests, and output regression benchmarks — against each monitored agent, scoring drift on a decay index with Slack/PagerDuty alerts; ships as a Docker container with a dashboard and webhook integrations.

Subset of the $40B+ observability market; AI-agent-specific monitoring is ~$2B and growing fast as enterprises move from single-model to multi-agent architectures.