Agent as a service
Agent as a service means a runtime runs your agents for you. You declare an agent and give it work through an API; the service handles sessions, sandboxes, tools, credentials, and recovery. OpenMA provides that runtime as open source: use it hosted, or run it on your own infrastructure.
With an agent SDK, the loop that calls the model and runs tools lives inside your process, and so does everything it needs: history, isolation, retries, and secrets. Agent as a service moves that loop into a runtime you call over an API. Your application creates a session, sends events, and streams progress back.
Because the agent is a service, the same agent can be called from your product, a scheduled job, Slack, a GitHub issue, or Linear. Vendor examples include Claude Managed Agents and the OpenAI Agents API; Moonshot describes Kimi Hosted Agents as harness as a service.
The unit of work. A session binds an agent version to an environment and owns an append-only event log, so work can run asynchronously, pause for input or approval, and resume later.
Tools run in a container or microVM, not in the API process. The service provisions the sandbox, applies network policy, and lets you choose the provider.
Shell, file, web, and browser tools out of the box, plus MCP servers and Skills you attach per agent.
Provider and tool credentials stay with the service and can be added to outbound requests, so tools in the sandbox do not hold them. In OpenMA, how strictly that is enforced depends on the sandbox provider and runtime profile.
Every input, output, and tool call is persisted before it is broadcast. If a harness crashes, a fresh one rebuilds context from the log and continues.
These are the same primitives used to compare runtimes on the AI agent platform page.
The question is where the agent loop runs and who operates it.
| Approach | Where the loop runs | What you give up or take on | Fits |
|---|---|---|---|
| Agent SDK in your app | Your process | You build persistence, isolation, retries, and credentials | One agent inside an existing service |
| Agent as a service (vendor) | The vendor's cloud | Model choice and where the control plane runs | Teams standardized on one model vendor |
| Agent as a service (open runtime) | Hosted by OpenMA, or your own infrastructure | You pick models, harnesses, and sandboxes; self-hosting means you operate it | Teams that want the service shape without lock-in |
Product-by-product comparisons are on the compare page.
Sign in at app.openma.ai, create an agent, and run a session from the Console, CLI, or API. Bring your own model key: hosted plans only charge for sandbox compute. See pricing.
Run the open-source runtime on Node.js, Docker, Fly.io, Vercel (Beta), or Cloudflare, with the database, queue, and object storage you choose. Self-hosting is free under Apache-2.0. See self-hosted AI agents and the deployment planner.
OpenMA exposes a Claude Managed Agents-compatible API and an OpenAI Agents API adapter over the same sessions, and lets you pick the model, harness, and sandbox independently.
A delivery model where a runtime runs your AI agents for you and exposes them through an API. You declare an agent (model, instructions, tools) and send it work; the service handles sessions, sandboxes for tool execution, credentials, persistence, and recovery.
An SDK is a library for the agent loop inside your own process, so persistence, isolation, and scaling are your job. Agent as a service moves that loop and its state into a managed runtime you call over an API. See Claude Agent SDK vs managed agents.
Closely related. An AI agent platform is the runtime itself; agent as a service describes consuming that runtime through an API instead of embedding an agent loop in your code.
No. OpenMA is Apache-2.0 licensed: use hosted openma.ai, or self-host on Node.js, Docker, Fly.io, Vercel (Beta), or Cloudflare.
Any configured provider: Anthropic, OpenAI, DeepSeek, OpenRouter, and custom endpoints through Model Cards. Hosted openma.ai is BYOK for the LLM.
The runtime primitives behind agent as a service.
OpenMA next to vendor runtimes, open-source runtimes, and coding agents.
Anthropic's hosted agent runtime, explained.
Run the service on your own infrastructure.
Moonshot's harness as a service, compared.
Self-host for free; hosted plans charge for sandbox compute.
