Last year, the team attempted to implement an LLM agent into enterprise workflows, but faced significant challenges, such as unpredictable behavior from minor wording changes, testing difficulties, and issues with traceability, which upset SOC-2 reviewers. To resolve these, they developed NOMOS, a step-based state-machine engine that offers predictability while leveraging the flexibility of any LLM (like OpenAI and Claude). Each state in NOMOS is explicit, testable, and can be independently managed, making it easy to version control. They have released the open-source core (MIT license) and provided documentation on their GitHub. Looking ahead, they are prototyping Kosmos, a platform akin to “Vercel for AI agents,” enabling the deployment of NOMOS and other frameworks under a unified control system. They invite others to join the waitlist for an exclusive membership and share their experiences with unreliable LLM agents.
Source link
