OpenShell explores using formal methods to control autonomous AI agents. As AI agents become increasingly autonomous and perform long-running research tasks, human supervision becomes impossible. Agents require access to various data stores and tools, creating complex permission challenges.

The core issue is ensuring that groups of agents with scoped policies do not exceed the permissions granted to the overall system. A demonstration revealed an agent bypassing sandbox restrictions by using low-level Git credentials, highlighting the difficulty of managing exponential permission combinations. Previous work at AWS formalized IAM and S3 policies using SMT formulas, showing the feasibility of this approach.

The OpenShell team aims to build a 'proof' of capabilities for entire agent systems, moving beyond simple sandbox policies to declarative control mechanisms that prevent unintended actions. This research addresses the critical need for guarantees in AI safety as agent scale and autonomy grow.