Cycles - Runtime authority for autonomous agents

Cycles provides runtime authority for autonomous agents, enabling you to control runaway agent spend and risky actions before they execute and cause damage.

It offers an open protocol with multi-language SDKs and is released under the Apache 2.0 license. Cycles is self-hosted and does not store prompts.

• Cost Control: Prevent agents from exceeding budget limits. For example, an agent that would have cost ~$10 in 12 seconds without Cycles can be capped at a $1 limit before executing its next action, preventing incidents that could cost thousands.

• Risk Mitigation: Gate actions, not just spend. Cycles can block risky actions, such as sending customer emails or performing mutations, before they are executed, even if the agent stays within its token budget.

• Blast Radius Isolation: Isolate every customer and prevent one agent's runaway actions from impacting others on the platform. Cycles ensures that internal actions proceed while sensitive external actions are blocked.

• Enforcement Evidence: Provide auditable proof of agent control. Cycles offers a pre-execution decision, scope, denial, commit, and authority trail, going beyond post-execution dashboards.

Cycles integrates with a wide range of AI stacks, including OpenAI, Anthropic, Spring AI, LangChain, LangGraph, Vercel AI, OpenClaw and many more. It offers quickstart guides for Python, TypeScript, Rust, and Spring Boot, with Docker deployment in under 5 minutes.

Key features include:

• Budget reservation before action execution.

• Action authority to control what agents do.

• Audit trail of all gated actions

• Multi-language SDKs (Python, TypeScript, Java/Spring).

• Integration with popular AI frameworks and providers.

• Tools like a Blast Radius Risk Calculator and a Cost Calculator.