AI agent platform
An AI agent platform is the runtime that turns a model into dependable work: it schedules sessions, runs tools in sandboxes, routes to models, keeps credentials out of reach, and recovers when something fails. Those four primitives are the basis for comparing platforms.
The platform sits between models and your products. You declare an agent (model, instructions, tools) and give it work. The platform orchestrates the rest: provisioning a sandbox, running the loop, persisting every step, delegating to other agents, and streaming progress. Open orchestration runtimes such as Google's AX and OpenMA, and vendor runtimes such as Claude Managed Agents and the OpenAI Agents API, all share this shape.
A chat window is one surface on top of a platform. The same agent might be called from Slack, a GitHub issue, a scheduled job, or your own app. Multi-agent orchestration, where a lead agent delegates to specialists that share or isolate files, happens in the platform.
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 after a crash or a deploy.
Where tools execute: shell, files, browsers, package installs. A platform isolates it from the API process, applies network policy, and lets you choose the provider (containers, microVMs, or a sandbox service).
The platform holds provider credentials, routes each agent to its configured model, and can add tool credentials to outbound requests so the sandbox does 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. A session runs in one place at a time, and if a harness crashes, a fresh one rebuilds context from the log and continues.
Anthropic calls this split decoupling the brain from the hands, explained in what Claude Managed Agents is.
Each layer does a different job and leaves different work to you.
| Layer | What it does | Who it's for | What it leaves to you |
|---|---|---|---|
| AI agent platform / runtime | Runs agents as a service: sessions, sandboxes, model routing, durability, API | Engineers integrating agents into products and workflows | You still write prompts, tools, and policies |
| Agent SDK / framework | A library for the agent loop inside your process | Developers building one agent or a custom stack | Operations: persistence, isolation, scaling |
| No-code / visual builder | A canvas to author agent workflows | Builders who prefer flows over code | A production runtime with real permissions and isolation |
| Chatbot UI | A conversation front end | End users | Agent execution, which a runtime behind the UI provides |
SDKs are covered in Claude Agent SDK vs managed agents and the OpenAI Agents API / SDK naming map. Coding agents such as Claude Code or OpenHands are a fourth category: tools developers drive, which can run inside a platform's sandbox.
Compared on the primitives above, from each vendor's documentation as of November 2026.
| Runtime | Type | Models | Self-host control plane | Sandbox |
|---|---|---|---|---|
| OpenMA | Open-source managed runtime | Any configured provider | Yes | Replaceable drivers |
| Claude Managed Agents | Vendor-managed runtime | Claude | No (self-hosted sandbox only) | Cloud or self-hosted sandbox |
| OpenAI Agents API | Vendor-managed runtime | OpenAI | No (self-hosted sandbox only) | Hosted or self-hosted sandbox |
| AX (Agent Executor) | Open orchestration runtime on Kubernetes | Model manifest | Yes, on Kubernetes with Agent Substrate | Sandboxed tasks on Agent Substrate |
OpenMA is an open-source AI agent platform: a meta-harness that manages sessions, sandboxes, tools, credentials, and recovery around the harness and model you choose. It exposes a Claude Managed Agents-compatible API and an OpenAI Agents API adapter over the same sessions, and delegates work to callable agents with their own histories.
You configure it through the Console, CLI, or API, and tools, policies, and custom harnesses are code. It is built for engineering teams running agents in production. It runs on your own infrastructure or hosted, and the source is open alongside other open-source AI agents.
The Claude Managed Agents alternative guide compares it with open and cloud options in one decision table.
Software that runs AI agents as a service. It manages sessions, sandboxes for tool execution, model and credential routing, and durable state, and exposes an API your applications call. It is the layer between the model and your product.
Coordinating agent work: scheduling sessions, running tools in sandboxes, routing to models, delegating to sub-agents, and recovering from failures. Multi-agent orchestration adds delegation between agents that share or isolate context.
No. A builder is an authoring surface. A platform is the runtime underneath. Some products bundle both. Production agents still need the runtime pieces: permissions, isolation, credentials, and durability.
No. OpenMA is configured through the Console, CLI, or API, and tools, policies, and custom harnesses are code. It is an orchestration runtime for engineering teams.
Tenant isolation, credentials kept out of the sandbox, an auditable event log, a choice of deployment region or self-hosting, and the ability to change model vendors without rewriting the platform.
The hosted runtime this platform category is often compared with.
Decision table for buyers evaluating Anthropic's hosted runtime.
Run the platform on your own infrastructure.
Open platforms, harnesses, frameworks, and coding agents.
When an SDK is enough and when you need a runtime.
Agents API vs Agents SDK vs AgentKit.
