README.md

September 2, 2026 · View on GitHub

Agent Kernel

The Operating System for Scalable & Compliant Enterprise AI Agents

Run, orchestrate, and deploy production AI agents at scale — across frameworks and clouds — without lock-in, rewrites, or fragile glue code.

📖 Docs • 🚀 Quick Start • ✨ Features • ☁️ Deploy • 💬 Discord • 🛠 Developer Guide

Why Agent Kernel?

Most agent frameworks help you build a prototype. Agent Kernel is the platform layer that gets you to production — with the governance, portability, and operational maturity that enterprises actually require.

	Agent Kernel
🔌 Framework-Agnostic	Run OpenAI Agents SDK, LangGraph, CrewAI, and Google ADK side by side. Swap with 2 import lines.
☁️ Cloud-Agnostic	The same agent code ships to AWS Lambda/ECS, Azure Functions/Container Apps, GCP Cloud Run, or on-prem.
🔁 Queue-Pipeline Execution	Every chat request runs through a queued pipeline: in-process by default (zero services, full retry/FIFO/dedup semantics locally), SQS, Kafka, and NATS JetStream transports for distributed deployments; a Helm chart ships the topology to any Kubernetes cluster.
🛡️ Compliant by Default	Built-in guardrails (OpenAI, AWS Bedrock), PII detection, full audit traces, jailbreak prevention.
🧠 Stateful & Knowledge-Aware	Pluggable session stores (Redis, Valkey, DynamoDB, Cosmos DB) + knowledge bases (ChromaDB, Neo4j, Starburst).
💬 Channels Built-In	Slack, WhatsApp, Teams, Telegram, Gmail, Messenger, Instagram — out of the box.
🔍 Production Observability	LangFuse, OpenLLMetry, and Pydantic Logfire tracing wired in. Every agent, tool, and LLM call — visible.
🤝 Open Standards	Native MCP (Model Context Protocol), A2A (Agent-to-Agent), and AG-UI (streamed event protocol for agent-facing frontends) support.
🆓 Apache 2.0	No licensing fees. No vendor lock-in. Production-ready open source.

⭐ If Agent Kernel is solving real problems for you, please star the repo — it's the single best way to help us grow.

🚀 Quick Start

Requirements: Python 3.12 – 3.13.x

pip install agentkernel

Build a multi-agent system that runs on OpenAI Agents SDK today and LangGraph tomorrow — same code:

from agentkernel.cli import CLI
from agentkernel.openai import OpenAIModule
from agents import Agent

math_agent = Agent(
    name="math",
    handoff_description="Specialist agent for math questions",
    instructions="You provide help with math problems.",
)

general_agent = Agent(
    name="general",
    handoff_description="Agent for general questions",
    instructions="You provide assistance with general queries.",
)

triage_agent = Agent(
    name="triage",
    instructions="You determine which agent to use based on the user's question.",
    handoffs=[general_agent, math_agent],
)

module = OpenAIModule([triage_agent, math_agent, general_agent])

if __name__ == "__main__":
    CLI.main()

That's it. Same code deploys to AWS Lambda, ECS, Azure Functions, or Container Apps with a single Terraform module. 👉 Get Started

✨ Features

🧩 Works with the Frameworks You Already Use

OpenAI Agents SDK • LangGraph • CrewAI • Google ADK • Smol Agents (soon) • LiveKit (soon)

No rewrites. No re-learning. Bring your agents — Agent Kernel handles the platform layer.

🛡️ Compliance & Governance — First-Class, Not an Afterthought

Enterprises can't ship agents they can't audit. Agent Kernel makes compliance the default:

Guardrails — OpenAI and AWS Bedrock guardrails for PII detection, jailbreak prevention, content moderation.
Pre/Post Execution Hooks — Inject policy checks, RAG context, redaction, or moderation around every agent call.
Full Traceability — Every agent action, tool call, and LLM invocation logged with configurable verbosity.
Observability — LangFuse, OpenLLMetry, and Pydantic Logfire tracing with a single config line.
Data Residency — Pick your cloud, your region, your storage backend. Your data stays where you need it.

📦 Sandboxed Code Execution

Let agents run code and shell commands in an isolated, permission-bounded environment — the platform handles it, your agent code stays clean.

Enable it in config — agents automatically gain code/command/file tools and the usage guidance is injected into their prompt.
Pluggable providers — local_subprocess (dev), docker (container-isolated), kubernetes (pod per sandbox, RBAC as the boundary), e2b (managed micro-VMs), daytona (cloud containers), and ec2_ssm (attach to an existing EC2 instance); bring your own via a dotted path.
Workload profiles — per-call, per-session, or shared lifetimes; each with its own permission policy (network egress, filesystem, CPU/memory, timeout) enforced fail-closed.
Per-user identity — run sandboxed code under the invoking user's identity, not one shared agent identity, via a pluggable principal resolver.

Learn more →

⏰ Deferred & Recurring Chats

Let a chat run later, or on a schedule — the platform owns the timers, the persistence, and the management API.

Enable it in config — a schedule block turns on deferring and the agent tools with no code change; mount ScheduleRESTRequestHandler when you also want the management routes.
One creation path, three callers — a schedule block on any chat request (acknowledged with HTTP 202), the agent's own create_schedule tool, or a direct ScheduleManager call.
Pluggable timers and stores — local (in-process, for development) or AWS EventBridge Scheduler for production; task records in memory, Redis, Valkey, or DynamoDB.
Managed over REST — list, read, amend, pause and cancel via /api/v1/schedules, scoped to the owning user by a pluggable Authoriser.

Learn more →

🧠 Memory, Sessions & Knowledge Bases

Layer	Backends
Session / Memory	In-memory, Redis, Valkey (AWS), DynamoDB (AWS), Cosmos DB (Azure), Firestore (GCP)
Conversation Threads	Persistent, named threads keyed by `session_id` — in-memory, Redis, Valkey, DynamoDB (AWS), Cosmos DB (Azure), Firestore (GCP)
Scheduled Tasks	Deferred and recurring chat execution — in-memory, Redis, Valkey, DynamoDB (AWS) task stores; local in-process or AWS EventBridge Scheduler timers
Vector Knowledge	ChromaDB
Graph Knowledge	Neo4j
SQL Analytics	Starburst Galaxy (Trino)
Custom	Pluggable `KnowledgeBase` interface — bring any backend

💬 Messaging Channels — Out of the Box

Slack • WhatsApp • Microsoft Teams • Telegram • Gmail • Messenger • Instagram

Build once. Ship to every channel your users live on. No bespoke bot code.

🔌 Open Protocols

MCP (Model Context Protocol) — Connect agents to external tools, data sources, and services. Optionally expose your agents as MCP tools.
A2A (Agent-to-Agent) — Native message passing, handoffs, and coordination between agents in a shared ecosystem.
AG-UI — Stream any agent's run (text, tool calls, reasoning, shared state) to a compliant AG-UI frontend, e.g. CopilotKit.

⚡ Built-In Execution Modes

Synchronous — Classic request/response.
WebSocket / Async — Streaming, long-running, real-time agent interactions.
Stateless MCP — Lightweight agent-as-tool deployments.

☁️ Deploy Anywhere

Same agent code. Pick your runtime. Full Terraform modules included.

Cloud	Serverless	Containerized
AWS	Lambda	ECS / Fargate
Azure	Functions	Container Apps
GCP	Cloud Run Serverless	Cloud Run Containerized
On-Prem / Kubernetes	✅ Docker image	Helm chart (baremetal + EKS, Kafka/NATS queue mode, KEDA autoscaling)

🛠 Agent Skills — Supercharge Your Coding Assistant

Install Agent Kernel skills and Claude, Copilot, Cursor, or Windsurf become experts at building production agents — no hallucinated APIs.

ak skill install

Skill	What it does
`ak-init`	Scaffold a new project — any framework, any deployment mode
`ak-build`	Add tools, agents, handoffs — context-aware and framework-specific
`ak-add-capabilities`	Wire in guardrails, tracing, sessions, MCP, A2A, AG-UI, hooks, multimodal, conversation threads, sandbox, scheduled tasks
`ak-add-integration`	Slack, WhatsApp, Messenger, Instagram, Telegram, Gmail
`ak-cloud-deploy`	AWS Lambda, ECS, Azure Functions, Container Apps, GCP Cloud Run with full Terraform
`ak-test`	Score, llm, and fallback test modes (pluggable evaluators) + a debugging playbook

See the use-cases/ directory for complete end-to-end examples built using these skills — each starting from a SPEC.md and generating a fully deployed agent.

🧪 Testing — pytest-Integrated

pytest tests/

Built-in score, llm, and fallback comparison modes, backed by a pluggable evaluator (DeepEval by default, or bring your own). CI/CD ready. Test agent behavior, not just code.

👥 Who's It For?

AI Startups — Go from prototype to production in days, not quarters.
Software Product Companies — Add intelligent agents to your SaaS without framework lock-in.
Software Services Companies — Ship client agent solutions in weeks, not months.
Domain Experts — Build production AI products without a full platform engineering team.
Regulated Enterprises — Get compliance, traceability, and data residency by default.

🌐 Affiliations

Agent Kernel is proud to be:

A member of the Linux Foundation
A member of the Agentic AI Foundation

💬 Community & Support

📖 Documentation: kernel.yaala.ai/docs
💬 Discord: Join the community
🐛 Issues: GitHub Issues
📦 PyPI: pip install agentkernel
☁️ Terraform Registry: Official modules
🛠 Developer Guide: DEVELOPER_GUIDE.md
🤖 Contributing with AI agents: AGENTS.md

License

Unless otherwise specified, all content — including all source and documentation files in this repository — is:

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

SPDX-License-Identifier: Apache-2.0