Agent Runtime

Run your agents as a secure, stateless service.

We build our agents, teams, and workflows with the Kern SDK and take them live using AgentOS. What does it mean to take an agent live? We should be able to:

Run the agent on demand, via an API request, or via interfaces like Slack, Telegram, and others.
Maintain long-running sessions that last from minutes to days to weeks.
Be durable across restarts, replicas, and infrastructure failures.
Be secure against unauthenticated access.
Monitor every run and action taken.

Kern's agent runtime, AgentOS, takes agents, teams, and workflows built with the Kern SDK or any other framework and runs them as a secure, scalable service.

Here's the smallest possible AgentOS example:

1from kern.agent import Agent
2from kern.db.postgres import PostgresDb
3from kern.models.openai import OpenAIResponses
4from kern.os import AgentOS
5
6db = PostgresDb(db_url="postgresql://user:pass@host:5432/kern")
7
8agent = Agent(model=OpenAIResponses(id="gpt-5.5"), db=db)
9
10agent_os = AgentOS(agents=[agent], db=db)
11app = agent_os.get_app()
12
13if __name__ == "__main__":
14    agent_os.serve(app="my_app:app", reload=False)

This script starts a FastAPI application that you can scale horizontally. Persistent sessions, streaming, JWT auth, tracing, and a scheduler, all come built-in.

What the runtime gives you

The runtime covers the ground between your agent code and a production service:

Concern	How AgentOS handles it
HTTP API	Auto-generated endpoints for every registered agent, team, and workflow
Persistence	Sessions and memory persisted to your `db`
Streaming	SSE on every run endpoint; tokens and tool calls stream as they happen
Auth	JWT validation with RBAC scopes built-in
Scheduling	In-process cron that polls the database and fires due jobs
Observability	OpenTelemetry tracing into the same `db`, queryable with SQL
Interfaces	Slack, Telegram, WhatsApp, A2A, AG-UI
Human in the loop	Pause runs for user confirmation, admin approval, or external execution