Claude Managed Agents: Runtime, Security Controls & Pricing

A hosted runtime, not a new model

Managed Agents is a set of composable APIs on the Claude Platform for building and deploying agents that run in Anthropic's own cloud rather than a customer's servers. It pairs an agent harness, the logic that decides when to call a tool, manage context and recover from errors, with the sandboxed execution, session persistence, credential management and permission controls that harness needs to run unattended. [1]

Sessions checkpoint and persist state, so an agent can run for hours and survive a disconnection. The Claude Console adds session tracing, integration analytics and troubleshooting views so a developer can inspect tool calls, agent decisions and failure modes after the fact. [1][2]

Atlas interpretation: None of these pieces are new in isolation; sandboxing, credential scoping and tracing all exist as products from other vendors. What changed is who runs them. A team that wanted a long-running agent used to have to build and operate that infrastructure itself. Managed Agents makes it a line item, billed alongside the tokens, rather than a system a customer's own engineers have to keep alive. [1]

Priced by the hour the agent is running

Standard Claude Platform token rates apply on top of a runtime fee of eight cents per session-hour of active operation, a consumption charge for the hosted infrastructure itself rather than for model output. Anthropic launched the feature as a public beta, reachable through the Claude Console, a CLI, or Claude Code using a bundled skill; multi-agent coordination and self-evaluation features were gated behind a separate request-access research preview. [1][2]

Named launch customers, in production

Anthropic named Notion, using it to delegate workspace tasks in parallel; Rakuten, for enterprise agents inside Slack and Teams across product, sales, marketing and finance; Asana, for AI teammates that collaborate inside a project; and Sentry, whose engineers used it to build debugging agents that write patches and open pull requests. Sentry's Indragie Karunaratne said the appeal was a runtime his team did not have to build: it gave them, in his words, "a secure, fully managed agent runtime," letting them focus on the handoff experience around it rather than the infrastructure underneath it. [1][2]

Anthropic reported internal testing on structured file generation showing task success rates improving by up to 10 percentage points over a standard prompting loop, with larger gains on more complex tasks. That figure is Anthropic's own internal benchmark, not an independently reproduced measurement. [2]

The model vendor becomes the hosting vendor

Atlas interpretation: Selling the execution environment alongside the model puts Anthropic in competition with the layer of startups and internal platform teams that had been building agent runtimes on top of the Claude API. It also moves more of an agent's operational surface, its credentials, its tool permissions, its logs, onto Anthropic's own infrastructure rather than a customer's. That is a natural extension of selling API access, and it is also a deeper dependency than calling a model endpoint. [1]