Agentic Layer

Organizations have successfully set up initial process automations with Generative AI. But now they face a new set of challenges: backend integrations are growing in complexity, long-term maintainability is uncertain, and compliance and security requirements keep increasing.

It is clear that agentic automation — where AI agents act autonomously to execute complex workflows — is the next logical step. But committing deeply to a rapidly evolving and volatile technology landscape carries real risk. How do you move forward without locking yourself into today’s tools and frameworks?

The Agentic Layer is an open-source, Kubernetes-native collection of building blocks for orchestrating multi-agent AI systems in a flexible, secure, and maintainable way. It acts as an abstraction layer and compliance guardian between your business applications and the fast-moving world of AI agents, models, and protocols.

It is composed of a set of spacers and controls that form an enterprise-grade platform with interchangeable components, adaptable to individual requirements.

The following picture visualizes how the Agentic Layer integrates between the frontends and applications on the top, the Kubernetes-based platform at the bottom, and your individual Agentic Workforces and Connectors that run your business.

There are two types of key spacers: The Agent Runtime and the different Gateways. The Agent Runtime takes care of running your agents and connecting them to other agents and tools / connectors. It allows for custom agent implementations and ready-to-use implementation templates, both with the option to use different frameworks, as long as they support the A2A protocol.

The Testbench connects to the Agent Runtime and makes use of Kubernetes to not only run your AI evaluations in the cluster, but also trigger the tests when the infrastructure (like agents and models) change.

The gateways control the traffic from and to the agents. They all are basically classical API gateways, but with a focus on the AI protocols. The Agent Gateway exposes Agents to the outside world, allowing for additional supported APIs, like making an agent compatible to a frontend that only supports OpenAIs Chat Completion or supports growing protocols like AG-UI or A2UI. Making an agent available through MCP is also an interesting solution until specific agent protocols gain more adoption.

The AI Gateway, sometimes referred to as model facade or model router, abstracts the call to the LLMs and allows switching providers quickly or hosting LLM models locally for certain use cases. It also allows for fine-grained control. Like adding guardrails that check for personal data, tracking LLM costs, or centrally log all interactions for compliance and auditing.

The Tool Gateway is the facade towards all the internal and external (MCP) tools. Besides beforementioned aspects, the authentication and authorization become especially important here, as this is where data is read and modified.

The Agentic Layer is designed to switch individual components easily and contributing new implementations independently. For each type of gateways, a lot of solutions, many of them open source, are popping up right now and are waiting for adoption. We highly appreciate contributions to support additional implementations.