Claude Managed Agents,
and the job they replace.
Managed Agents entered public beta on 8 April 2026 as a fully hosted agent harness, sandboxing, tools and event streaming handled by Anthropic rather than assembled by you. Four months and several updates later, the useful question is not what it can do. It is which of your tasks are actually agent-shaped.
Managed Agents is Anthropic’s fully hosted agent platform: a way to run multi-step, tool-using Claude workflows with sandboxing, memory, orchestration and webhooks handled server-side rather than assembled from scratch. It sits below the surface of some Cowork-style experiences, but it is a distinct, developer-configured layer, not a rebrand of Cowork itself.
A hosted harness, not a framework you assemble
Managed Agents launched on 8 April 2026 alongside ant, a dedicated command-line client for the Claude API. Sandboxing, tool access and event streaming are handled by Anthropic’s platform rather than built by you, which is the actual shift: the barrier to running a multi-step, tool-using agent with limited supervision dropped from “build an orchestration layer” to “configure one that already exists.”
Memory and multiagent orchestration arrived within a month
Memory support for agents reached public beta on 23 April, letting an agent retain context across a session rather than starting cold. Multiagent orchestration and a companion feature called Outcomes followed on 6 May, letting several agents divide work on the same problem rather than one agent doing it all in sequence.
The July additions: effort, webhooks and session seeding
On 22 July, three further changes: effort level configuration per agent, an expanded set of webhook event types covering environment and memory store changes, and the ability to seed a new session with up to fifty prior events so an agent starts with real context instead of nothing.
The design question that actually matters
Multiagent orchestration is genuinely useful for work that decomposes cleanly, research fanned out across sources, a build-then-verify pipeline, and genuinely unhelpful for a task that needs one consistent line of judgement throughout. Getting that distinction right before configuring anything is most of the value; skipping it is the most common mistake.
Start with the smallest reliable workflow
The steepest part of the learning curve is not the API, it is picking a first task you already understand well enough to judge whether the agent actually got it right. Pick that before picking the hardest, least-understood problem you have.
No, and it is worth being precise about the difference. Claude Cowork is the product experience you brief and receive a deliverable from. Managed Agents is the API-level platform some agentic workflows, including some Cowork-style ones, run on underneath. If you do not code day to day, understanding that distinction is exactly the kind of thing a Claude Cowork training session covers before anything gets built.
Things people
usually ask.
No. Cowork is the product experience you brief and read results from. Managed Agents is the underlying platform primitive developers configure directly, sandboxing, tools and event streaming as a hosted service. Some Cowork-style workflows now run on it, but the two are not interchangeable.
Mostly yes. It is configured through the API or the ant command-line client, not a chat window. If you do not code day to day, Claude Code for non-developers is a better starting point for understanding what it can do before a developer builds anything.
On 22 July 2026, effort level configuration per agent, an expanded set of webhook event types for environment and memory store changes, and the ability to seed a session with up to fifty prior events.