# Cloudflare details evidence-first agents for Managed Defense early beta > Cloudflare’s Managed Defense agent harness collects evidence before model analysis, with analysts responsible for decisions and mitigation. Canonical URL: https://www.devobs.io/news/news-event-0b5f1f77b7190e9348ff943f/ By: Nina Patel Published: 2026-10-08T05:37:30.666Z Updated: 2026-10-08T05:37:30.666Z Event date: 2026-10-07 Section: Security Cloudflare says an evidence-grounded, multi-agent investigation harness is available in **early beta** within Managed Defense for eligible application-security alerts and cases. It assembles evidence and recommendations; Managed Defense analysts remain responsible for decisions and mitigation, [according to Cloudflare](https://blog.cloudflare.com/agentic-security-operations/). ## Collection before inference Cloudflare describes fixed reconnaissance workflows that collect customer context, detection history, traffic baselines, enforcement outcomes and network observations before model analysis begins. Each evidence item carries source, version and timestamp information. Application code fixes customer scope before any model sees results. Clef, Cloudflare’s decision model running on Workers AI, performs initial triage. Alerts scored as highly likely to be false positives skip specialist analysis. For alerts needing deeper review, a coordinator runs four specialists in parallel: traffic analysis, customer context, global telemetry and threat intelligence. The threat-intelligence specialist checks indicators already admitted to the investigation. A synthesis agent combines the findings without fetching new evidence. These are components of [Cloudflare’s described architecture](https://blog.cloudflare.com/agentic-security-operations/), not independent performance findings. ## Why the evidence boundary matters The engineering distinction is that evidence collection and customer-scope enforcement sit in application code rather than relying on model prompts. Cloudflare says the same reconnaissance snapshot can be replayed, allowing comparisons between specialist interpretations without changing their inputs. That makes reproducibility a useful property of the design, but reproducibility alone does not establish detection accuracy. One important evaluation target is the triage gate: how often does it route a real incident away from specialist review? Cloudflare also says the advisory distinguishes **not checked**, **checked with no matching result**, and **checked with evidence supporting absence**. When evidence is insufficient, the system makes no classification or disposition recommendation. For example, unavailable global telemetry limits an advisory to describing what is unusual for the customer, rather than whether the pattern is widespread. [These distinctions](https://blog.cloudflare.com/agentic-security-operations/) give evaluators another concrete check: whether missing evidence remains visible in the final recommendation. ## Availability and roadmap The announcement describes an early beta, not general availability. For fully managed customers, Cloudflare says analysts can apply suggested rules; other customers will receive recommendations through the dashboard and their chosen alert path. Cloudflare also plans a more flexible Custom Managed level and says it will explore continuous agents that monitor traffic for patterns. Those plans are separate from the early-beta capabilities described in [the announcement](https://blog.cloudflare.com/agentic-security-operations/). ## Source references -