Definition
What managed agents are
Managed agents are AI agents that run on infrastructure a vendor operates for you: the vendor hosts the agent loop, sessions, and state, and your product talks to them through an API.
In plain terms: an agent needs more than a model. It needs a loop that plans and calls tools, a sandbox where files and commands run, sessions that survive pauses, and a way to stream progress and results back to whatever asked for the work. A managed-agent service packages that operational layer behind an API, so your product submits work and receives events instead of running agent infrastructure.
The defining line is operational responsibility. With a framework or a bare model API, you assemble and operate the runtime. With managed agents, a provider operates at least the agent execution layer for you. Where the sandbox runs varies by product: some run it entirely on vendor infrastructure, some let the sandbox sit on infrastructure you control while the loop and state stay managed.
Anatomy
What a managed-agent service actually manages
Anthropic's documentation for Claude Managed Agents describes the category's anatomy well, in four concepts that make a useful map for the wider category:
- The agent: the configuration, a model, a system prompt, tools, MCP servers, and skills, defined once and referenced across runs.
- The environment: where execution happens, a sandbox with file, shell, and network access, hosted by the vendor or, in some offerings, on your own infrastructure.
- The session: a running instance of the agent doing a specific task, holding state that can pause and resume.
- Events and outputs: the streamed messages, files, and artifacts your application consumes, with history the service persists.
Not every product manages every layer, and that is the first evaluation question: which of these four does the vendor operate, and which remain yours.
Lifecycle
How a task runs through a managed agent
A typical lifecycle looks like this. You define the agent, or pick a prebuilt one. Your product calls an API to start a session with the task. The service provisions an execution environment, runs the agent loop inside it, and streams progress back; what the events carry varies by vendor, tool calls, file changes, and status updates among them. If the task pauses, the session and its state persist server-side; when it resumes, the agent picks up where it stopped. When the work completes, your product collects the outputs, files, artifacts, and a record of what happened.
What you did not do in that story is the point: no loop code, no session store, no streaming plumbing, and in the vendor-hosted default, no sandbox provisioning either. That is the burden the category removes.
Boundaries
Managed agents vs frameworks, model APIs, and serverless
An agent framework, LangGraph or CrewAI for example, helps you define agent behavior in code you then have to run somewhere. A managed-agent service operates the running. The categories meet in the middle: AWS's AgentCore, for instance, operates agents you build with external frameworks.
A model API generates responses; it does not execute work in an environment. Models generate tokens. Harnesses complete work. Managed agents sit on the harness side of that line: what is being managed is the working system around the model, not the model alone.
Serverless describes an execution and scaling model, not who operates the infrastructure. Some managed-agent services are also serverless, provisioning per session or per task and billing for active time, but serverless is a property some products have, not the category's definition.
Landscape
Three shapes of managed agents
The market splits into three shapes. Vendor-native stacks: Anthropic's Claude Managed Agents, in public beta since April 2026, runs Anthropic's harness with Claude models; Google's managed agents in the Gemini API, previewed at I/O 2026, provision Google's agent in a hosted sandbox. Each manages its own vendor's stack, deeply and only.
Bring-your-own-agent infrastructure: Amazon Bedrock AgentCore, generally available since October 2025, operates the runtime for agents you build, with an optional managed loop of its own. You keep the freedom of your own agent code and keep most of the responsibility for what the agent does.
Multi-harness platforms manage existing harnesses, the complete agents the labs and open-source communities already ship, and make the harness a choice instead of a commitment. Which shape fits depends on how committed your product is to one vendor's stack; for the vendor-by-vendor scoring, see our managed agents comparison.
Evaluation
What to check before you commit
Managed agents trade infrastructure burden for coupling. With single-vendor offerings, sessions, state, and permission formats follow that vendor; the questions worth asking before committing a backend are about that coupling.
- Which layers move to the vendor, and which stay yours: the loop, the sandbox, the state, or all three.
- Whether the sandbox can run on your infrastructure when compliance requires it, and what still lives on the vendor's side when it does.
- Whether the contract is open or proprietary: can another runtime speak it, or does leaving mean rewriting.
- Whether you can compare agents on your own tasks, or the choice of agent is fixed by the platform.
- How the meter runs: token costs, active runtime, session time, and what idle costs.
Across vendors
Managed agents across harnesses
HarnessRouter is the world's first unified interface for agent harnesses, and it applies the managed-agents pattern across vendors: one API that operates eight harnesses, Codex, Claude Code, Hermes, Pi, DeepSeek Harness, OpenCode, Qwen Code, and Cline, in per-run sandboxes, with sessions, streamed task, run, session, file, and artifact events, and recorded traces.
The contract underneath is the open Unified Harness Protocol, so the integration is not proprietary to us, and the Apache-2.0 Community Edition runs the same eight harnesses on your own infrastructure, which answers the self-hosting question with more than a sandbox. Execution is serverless per task: a sandbox exists while your task runs, and nothing of yours runs between tasks.
Under this article's definition, that is managed agents with one difference in kind: which agent does the work stays a parameter on the request, so the choice can follow the evidence rather than the platform.
FAQ
Managed agents FAQ
What are managed agents in one sentence?
Managed agents are AI agents whose operating infrastructure a provider runs for you: the provider hosts the agent loop, sessions, and state, and your product submits work and receives results through an API.
What is the difference between a managed agent and an agent framework?
A framework helps you define agent behavior in code you still have to run and operate. A managed-agent service operates the execution layer: the loop, the sandbox, sessions, and event delivery. The two can meet: some managed platforms operate agents you build with a framework.
Are managed agents the same as serverless agents?
No. Managed describes who operates the infrastructure; serverless describes how it provisions and scales. Some managed-agent services are also serverless, and HarnessRouter, for example, scales to zero per task, but a managed agent could run on always-on infrastructure and still be managed.
How are managed agents priced?
Meters vary by vendor and change often: model tokens, active runtime, session hours, or per-second compute, sometimes combined. The stable question to ask is what idle time costs. Check current vendor pricing pages rather than any summary, ours included.
Can managed agents run on my own infrastructure?
Sometimes, and the phrase self-hosted can mean different layers. Some vendors let the sandbox run on your infrastructure while the loop and session state stay managed on theirs. HarnessRouter's open-source Community Edition moves the whole layer: the same API, sandboxes, and harnesses on infrastructure you control.
Which companies offer managed agents?
As of August 2026: Anthropic (Claude Managed Agents, beta), Google (managed agents in the Gemini API, preview), AWS (Amazon Bedrock AgentCore), and HarnessRouter, which manages eight harnesses across vendors behind one API. See our managed agents comparison for the vendor-by-vendor detail.
The managed pattern, with the choice kept open.
Run Codex, Claude Code, Hermes, Pi, DeepSeek Harness, OpenCode, Qwen Code, and Cline as managed agent harnesses through one API.
Start building free