Platform primitives for trustworthy AI agents
rossoctl is a set of platform primitives for agent security, resilience, and efficiency that platform owners can build on. It comes in two parts: RossoCortex, a data plane that mediates what agents do, and a set of services agents can call. It's open source, framework-neutral, and built on open standards, with A2A between agents and MCP for tools.
The problem
AI agents don't behave like the cloud applications we know how to run. They decide which tools to call while they run, they don't cleanly separate data from instructions, their goals drift as context changes, and they can't reliably report what they did. Those traits make the ordinary demands of production hard to meet: failure recovery, admission control, security, and resource management.
Agent systems fail in ways enterprise architecture wasn't built for, so they need a platform to supply the guarantees they can't supply themselves. Kubernetes solved the same kind of problem for cloud software by separating application logic from placement, scaling, and recovery. An agent platform needs an analogous split: decouple agent logic from the guarantees around admission, isolation, and failure recovery that production demands.
RossoCortex, the data plane
RossoCortex is an intercept. It sits transparently between an agent and everything external it touches: models, tools, users, and other agents. From that one vantage point it enforces the guarantees agents can't enforce for themselves. It is flexible by design: it works with different agent types, including black-box harnesses, integrates with different sandboxes, and works across container network choices.
Rossoctl services
The services are the building blocks agents need to do real work: skills, tools, memory, knowledge bases, and sandboxes. The focus is on capabilities that lower the risk of agentic workloads and use resources more efficiently.