✦Architecture

One runtime between your channels and your agents

Requests arrive from any channel or protocol. Agent Kernel runs your framework-native agents with sessions, guardrails, queues and sandboxes, on whichever cloud you choose.

Sources · Channels & clients

Messaging channels

Protocols & APIs

Apps, voice & schedules

Agent Kernel · Runtime

Agent Kernel

Runtime

Sessions & Memory Hooks & Guardrails Sandbox & Scheduling Queue Pipeline & Scaling

Destinations · Agents, data & clouds

Agent frameworks

Memory & knowledge

Clouds & observability

Request · Inbound

BuildConnectRememberGuardScaleObserve

Response · Outbound

✦Built for Every Level of Expertise

Agent Kernel, Explained For Your Level of Expertise

Pick the path that fits you best

No code required

Business Leader

You run or work in a business/enterprise and want to incorporate AI agents that actually work into your business workflows without needing to handle the complexities of the tech.

Choose this path

Some coding

Developer

You build software but haven't built AI agents yet. You want to ship something robust and real without learning a new stack from scratch.

Choose this path

Deep tech

AI Engineer

You already work with LLMs and agentic frameworks. You need a production-grade AI agent execution framework that doesn't get in your way.

Choose this path

✦Features

Everything you need for production agents. In one runtime.

One runtime for frameworks, channels, memory, guardrails, sandboxes and deployment, with the observability and testing an enterprise team actually trusts.

Build

Write agents in the framework you already know. Agent Kernel wraps them with one API for tools, hooks, sessions and structured output, so the agent logic stays yours.

Framework Adapters

OpenAI Agents SDK, LangGraph, CrewAI, Google ADK, Smolagents and Pydantic AI run side by side in one runtime. Switch frameworks by changing two import lines.

OpenAILangGraphCrewAIADKSmolagentsPydantic AI

Docs Example

Portable Tools

Write a tool once as a plain Python function and bind it to any framework through ToolBuilder. The tool context exposes the runtime, agent, session and request.

ToolBuilderToolContextAny framework

Docs

Execution Hooks

Pre- and post-hooks wrap every run: inject context, validate input, rewrite or halt replies, and filter streamed events as they are produced.

Pre-hooksPost-hooksStream events

Docs Example

Structured Output

Return typed JSON instead of prose. Pydantic schemas map to each framework's native structured-output mode and come back as one reply type.

PydanticJSONAll frameworks

Docs Example

Multimodal Input

Images and files ride along with text. A vision model describes each attachment, stores it outside the session, and an analyze tool pulls details on demand.

ImagesFilesVision LLM

Docs Example

Multi-Agent Systems

Compose specialist agents, even from different frameworks, in one runtime with handoffs and a shared session.

HandoffsMulti-frameworkShared session

Docs Example

✦Deployment

Deploy Anywhere

Run the same agent code on AWS, Azure, GCP, or your own Kubernetes cluster. Zero rewrites.
Includes production-ready Terraform modules and a Helm chart with best practices baked in.