Agent Harnesses
Agent Harnesses
Run agents built with your preferred harness in your own infrastructure.
AgentOS can run agents built with any framework or harness. It provides the service layer around them, handling orchestration, authentication, streaming, session storage, and background execution.
Use Claude or Codex to power agents inside your product. Connect your tools, give your agent instructions on how to execute those tools, and let your application submit requests through the AgentOS API.
The harness manages the model and tool loop; AgentOS manages the harness.
For example, a support assistant can look up a ticket, check an order through your MCP server, and prepare a resolution. Your application streams the response to the user and keeps the session available for follow-up questions. The same service can run longer tasks in the background.
Start with your harness
Run Claude
Run Codex
Both guides follow the same pattern so you can see what changes when you choose a different harness.
What AgentOS handles
AgentOS handles everything around the harness: orchestration, authentication, streaming, session storage, and background execution. It makes the harness available to your application as an API and MCP server.
| Responsibility | What you configure with AgentOS |
|---|---|
| API and run orchestration | Register agents, accept requests, start background work and manage cancellation. |
| Authentication | Protect your service and authorize callers using AgentOS security configuration. |
| Streaming | Deliver text and tool activity to your application as server-sent events. |
| Sessions and run history | Store conversations, run status and captured tool calls and results in your database. |
| Background execution | Let work continue after the client disconnects; configure a durable queue for supported jobs that must survive worker failure. |
| Multiple replicas | Use shared storage and coordination so requests can reach different API replicas. Native conversation recovery depends on the harness. |
Keep sessions across requests
Use a stable session ID to continue a conversation. Agno run history and the harness's native conversation state are separate parts of that session:
- Claude can mirror native transcripts through Agno's database-backed session store and load them on another replica. The SDK still uses local files during execution.
- Codex stores its native thread identity with the Agno session. When that native thread is unavailable, the current adapter can start a new thread using bounded Agno history. This is history recovery rather than exact native continuation.
Conversation recovery does not restore files the agent changed or processes it started. See Sessions and State for storage, concurrency and recovery behavior before deploying multiple replicas.
Build toward deployment
Once your first service is running:
- Connect Your Tools and configure their permissions.
- Configure Your Agent with instructions and native SDK options.
- Sessions and State explains what persists between requests.
- Background Runs and Streaming covers polling, reconnecting and cancellation.
- Deploy and Operate covers authentication, shared storage, retries and recovery.
AgentOS also supports LangGraph, DSPy and Antigravity integrations. Explore Other Integrations when your agent uses one of those runtimes.