Cloudflare has detailed how its Managed Defense team applies a harness of specialized AI agents to security alert analysis. The system runs on Workers and draws on global network telemetry, and its defining design choice is a hard separation between deterministic evidence collection and model inference. Rather than letting a model reason over raw, unverified alert data, the harness first gathers concrete evidence, then feeds that grounded context into inference. The result is recommendations delivered to Managed Defense Analysts that trace back to verifiable inputs instead of model speculation.
The concrete architectural changes described are:
- A team of specialized AI agents rather than a single general-purpose model, each scoped to a role within the analysis workflow.
- The agents are built on Workers, Cloudflare's serverless runtime, so compute runs on the same global network that produces the telemetry.
- Global network telemetry is used as the primary data source for analyzing security alerts.
- Deterministic evidence collection is split from model inference, so evidence is gathered by predictable, reproducible steps before any model is invoked.
- Recommendations are surfaced to Managed Defense Analysts, keeping humans in the review loop rather than automating response end to end.
- The system is framed as a harness: the agents operate inside structured constraints rather than acting autonomously.
For developers building agentic systems, the pattern here is portable. Grounding an agent by separating evidence gathering from inference directly addresses the hallucination and unverifiable-output problems that plague LLM-driven tooling. Running that harness on Workers shows how edge compute plus first-party telemetry can keep latency and data locality under control. The emphasis on specialized agents over one monolithic model is also a practical argument for decomposition: narrower agents with defined inputs are easier to test, audit, and constrain. Anyone wiring LLM agents into production workflows—especially security or observability tooling—can borrow this evidence-first structure regardless of platform.
Source: https://blog.cloudflare.com/agentic-security-operations/.