Runtime Controls: Leave Your Agents Running Unattended
Serverless GPU Functions, agent SDKs, hardware-isolated sandboxes, and modern AI models make it more practical than ever to run autonomous agents in production. Yet for many teams, the “leave your agent running overnight” dream still feels out of reach. The hesitation is not about model intelligence. It is about predictability when the agent makes a bad decision. Anthropic’s recent research reports that among the longest-running sessions, the length of time Claude Code works before stopping increased from under 25 minutes to over 45 minutes within three months. We interpret this as operators showing greater trust in agents running more autonomously, with oversight shifting to higher-level supervision. Even so, usage remains concentrated in software engineering, while higher-stakes domains, where actions are harder to reverse, still represent a small share of usage. As agents move into these domains, the constraint shifts from capability to operational trust. Organizations are less concerned with whether an agent can act and more concerned with what happens when it makes a mistake.Introducing Runtime Controls 🪐
To make longer-running autonomy more practical, we launched Runtime Controls, a guardrail layer around functions and tool calls during agent execution. Runtime Controls help keep agents bounded while they operate:- gate risky actions before execution
- prevent loops and runaway tool usage
- enforce call budgets and policy gates
- add circuit breakers, retries, and cancellation
- provide runtime observability