CTRLRun

•
28
•

Establishes a critical execution safety layer for AI agents, acting as the last check before irreversible actions.

CTRLRun is a Python library designed to provide an execution safety layer for AI agents, ensuring that consequential actions are performed at most once, exactly as approved, and with a transparent audit trail. It addresses critical issues like duplicate effects from lost responses, unauthorized action mutations, concurrent agent conflicts, and authority escalations by sitting between an agent's decision to act and the actual tool call. By normalizing, deciding, approving, reserving, executing, and recording every protected action, CTRLRun prevents common agent pitfalls and introduces a human-in-the-loop mechanism for ambiguous or high-risk operations.

主な機能

01Binds human approvals to the exact action hash proposed, preventing mutation or replay attacks.
02Generates portable, auditable JSON receipts for every protected action, chained to its policy hash.
031 GitHub stars
04Enforces concurrency control by atomically reserving effect keys across distributed processes and hosts.
05Guarantees exactly-once execution for consequential actions, handling ambiguous outcomes from lost responses.
06Provides a flexible, YAML-based policy engine to define autonomous, human-approved, or denied actions.

ユースケース

01Enforcing granular permissions and delegated authority for agents, preventing escalation or unauthorized access to sensitive systems.
02Preventing duplicate or mutated financial transactions, such as refunds, by agent workflows.
03Ensuring safe and controlled execution of infrastructure operations, like Kubernetes namespace deletions, with proper approvals and concurrency checks.