Cloudflare Agents SDK: Durable Identity, State and Scheduling

Evidence check: 25 October 2026. Product details reflect the referenced documentation observed at this cutoff; exact plans, prices, regional eligibility and unverified release dates are not claimed. This is an editorial guide, not a hands-on test.

Runtime versus harness

Cloudflare’s October 2026 docs distinguish the Agents SDK runtime from the harness: the runtime handles agent identity, state, connections, scheduling and recovery; the harness organizes model calls, context and tool selection. The distinction matters because some workloads need durable background logic without an autonomous model planning loop.

Stateful execution

The SDK documentation describes an Agent class backed by Durable Objects with SQLite state, scheduling APIs and real-time connections. The long-running guide explains that persisted state survives hibernation, while in-memory fields, timers and open requests do not. Building on the wrong persistence assumption can lose a task that appeared successful before a worker was evicted.

Where Workflows belongs

An agent can schedule itself to check conditions, while an independent multi-step pipeline may belong in Cloudflare Workflows. Use the latter for operations requiring separate retries, approval and longer recovery semantics. Keep a stable job ID and external-effect ledger regardless of mechanism. Long-lived identity is not permission to perform unattended writes indefinitely.

Security decisions

Design per-user authorization and tenant isolation before exposing callable agent methods. Restrict outbound destinations and connector credentials. Validate an inbound webhook rather than letting it invoke an arbitrary tool. Keep model input and retrieved web content lower-trust than user-approved policy. Logs may contain sensitive payloads, so configure redaction and retention.

Practical scenario

A monitoring agent schedules a periodic check, reads an allowlisted endpoint and updates durable state. When a meaningful condition occurs, it creates an approval request or notification event. A separate workflow performs any write. Test sleep/wake recovery, duplicate triggers, status checks and behavior during a temporary outage.

Dates and procurement

Cloudflare’s referenced documentation has explicit August and October 2026 update dates, but those are documentation dates rather than a universal SDK release date. Pricing, quotas and regional entitlement are omitted until verified for the selected deployment. Compare a queue-driven worker or a standard workflow engine when no conversation state is needed.

Keep scheduling and effect execution separate

A durable agent might wake every morning, retrieve an approved set of metrics, generate a summary and store a reviewable draft. Persist a unique run ID and target period so that repeated alarms or an interrupted process cannot create duplicates. Require approval before external notifications or data changes. Cloudflare’s runtime and scheduling docs explain available building blocks; they do not guarantee every endpoint or plan has equal access.

Operating checklist

Record agent namespace, identity boundaries, storage keys, schedule definition, clock zone and deployment version. Exercise pause/resume behavior during deploys and concurrent invocations. Inspect durable state for sensitive payloads and define deletion/retention policies. Identify whether a specific operation depends on Workers, Durable Objects or another billed service, and obtain effective-dated prices separately. Test a network failure after an API submission and reconcile actual remote state before retrying. A durable runtime keeps context available; it is not automatic transactional consistency with outside systems.