An agent needs more than a model: a place to run, memory, tools and clear limits on what it can do. Cloudflare brings those pieces together, so you can build the application and put controls around it on the same platform.
Build agents that do useful work
State, without the servers. The Agents SDK runs on Durable Objects, with persistent state, SQL storage and scheduled tasks. Your agent can remember work between requests without you operating a server.
Tools beyond chat. Connect your agent to documents with AI Search, let it browse with Browser Run, or use Workflows for multi-step jobs that can retry and resume.
Your choice of model. Run open-source models on Workers AI or connect to external model providers through AI Gateway. Choose the model that fits the task.
Protect the application. Combine WAF rules with AI Security for Apps detections to identify and block malicious prompts, including prompt-injection attempts, at your AI endpoints.
Control model traffic. Route model calls through AI Gateway to see usage and costs, set rate and spend limits, and apply guardrails to prompts and responses.
Limit access to tools. Use Access policies and curated MCP portals to control who can reach which tools. Add human approval in your agent code before consequential actions run.
Build: download the event starter and deploy it as a Worker in your team’s own Cloudflare account. Then give it one tool that does something real for your industry, memory, one platform primitive that makes it more than a chat wrapper, and a human approval for one consequential action. Secure it (optional, nothing scored): explore AI Gateway, AI Security for Apps + WAF and Secure MCP. Submit a project name, demo URL and repo link; staff judges score the live demos.