agent-native-architecture

Build edge agents by defining outcomes in prompts rather than code, leveraging Cloudflare primitives like Durable Objects, Queues, and KV. The agent interprets the prompt to decide which tools to use, reducing tight coupling to implementation details and enabling rapid iteration.

Yes. Prompt-native agents access Cloudflare primitives—durable_object_rpc, queue_send, kv_put, service_binding_call—as tools the agent decides how to deploy. Outcomes and workflows emerge from prompt specification, not hardcoded logic.

Prompt-defined agents specify desired outcomes in prompts and let the agent choose edge primitives to achieve them, enabling self-modifying systems and faster iteration. Code-defined architectures lock behavior into implementation, creating tighter coupling and slower adaptation cycles.

Start with outcome specification in the prompt, iterate rapidly using edge primitives, then graduate mature logic to code when patterns stabilize. Context management strategies and verification patterns guide scaling stateful agents across Cloudflare's edge infrastructure.

No. Prompt-native architecture lets agents decide tool usage through prompts alone. Start with a prompt describing the desired capability—such as 'create a stateful chat agent using Durable Objects'—and let the agent design the flow.

The philosophy requires well-formed prompts and clear primitive bindings; agents must have sufficient context about available tools. Verification patterns and gradual code graduation prevent drift as complexity increases and reliability requirements rise.