Managed Agents

Anthropic launched Claude Managed Agents, a cloud-hosted runtime that lets developers build and deploy production agents without owning the infrastructure underneath.

Same week OpenAI shipped workspace agents. The managed agent category coalesced across both labs in a week.

Why I care about this one

I build agents at work. Every one of them needs the same boring stuff: a sandbox, authentication, tool execution, state management, error recovery. The boring stuff is where 70% of the build time goes and none of the differentiation lives.

A managed runtime takes that work off the table.

What it is

Claude Managed Agents is a cloud platform. You define the task, the tools the agent can use, and the guardrails. Anthropic runs everything else: orchestration, context management, sandboxing, retry logic, execution tracing.

Long-running sessions are built in. An agent can run autonomously for hours, hold persistent state, and pick up where it left off. Multi-agent coordination lets one agent spawn and direct others, still in research preview.

Pricing is consumption-based. Token rates match the standard Claude Platform. Active session runtime bills at $0.08 per session-hour on top.

How to use it

Four entry points depending on how you already work.

Claude Console. Web interface. Configure, deploy, and monitor agents without touching a CLI. Best for first-time setup and non-engineers.

New CLI tool. Deploy agents from your terminal. Pairs cleanly with existing CI/CD. Best for teams already shipping infrastructure-as-code.

Claude Code with the claude-api skill. Build and deploy agents from your editor using the same patterns you already use for coding. Best for developers who live in Claude Code.

Partner integrations. Notion, Asana, Slack, Teams all have first-class hooks. Agents plug directly into the tools where work already happens.

Scoped permissions, identity management, and execution tracing ship by default. An agent can only touch what you let it touch, and every action is logged.

Where it’s strong

Removes the boring infrastructure layer. Sandboxing, auth, retries, state, tracing: all of it for free. What used to take a week of DevOps now takes an afternoon.

Long-running autonomy. Agents that run for hours without re-prompting, holding state through the session. Research an account overnight. Draft a report over lunch. Monitor a channel continuously.

Enterprise-ready out of the gate. Identity, permissions, tracing, audit logs ship with the platform. Anyone who has tried to get a homegrown agent past IT knows how much work that saves.

First-party partner hooks. Notion, Asana, Slack, Teams integrations mean you are not building connectors from scratch. Pick your workflow, plug in.

Where it falls short

Pricing compounds fast. Token costs plus $0.08 per session-hour. An agent that runs 12 hours on a heavy session is a line item worth modelling before you deploy a swarm.

Multi-agent coordination is preview only. The feature that unlocks the bigger pattern is gated behind a research preview access request. For now you are building mostly single-agent workflows.

Vendor lock-in by design. Managed runtimes are sticky. You write against Anthropic’s SDK, deploy to Anthropic’s infrastructure, pay Anthropic’s runtime bill. Nothing wrong with that, but it is not portable.

Debugging is different. When an agent fails in a managed runtime, you are reading execution traces instead of stepping through code. The observability is strong. The debug loop is still new for most teams.

My take

Every major lab now has a production agent platform. OpenAI shipped workspace agents earlier this week. Anthropic shipped managed agents today. Google will ship something similar inside a quarter.

The category collapsed in a week.

What that means practically: you no longer need to build agent infrastructure. You need to build good agents on top of someone else’s runtime. The boring stuff is commoditised. Your edge is the task design, the tool selection, the guardrails, and the feedback loops.

If your team still builds its own sandbox, retry logic, and state management, stop. That work is now someone else’s SaaS. Spend the time on what the agent does, how it decides, and how it gets better.

The runtime will keep changing. Whether you use Claude, OpenAI, or something else a year from now, the pattern is the same. Write the task, define the tools, set the guardrails, trace the outcomes.

Agents as a platform is settled. Agent design is the work now.

- JC