Anthropic Managed Agents: Full Guide to Hosted Agent Infrastructure

What Managed Agents Is

On April 8, 2026, Anthropic launched Claude Managed Agents in public beta — a fully managed runtime for building and deploying autonomous AI agents. Instead of building your own agent loop, tool execution layer, sandbox, and orchestration logic, Anthropic provides all of that infrastructure as a service.

Early adopters include Notion, Asana, Rakuten, Sentry, and Netflix.

Architecture and Key Features

Managed Agents uses four primitives: Agents (reusable configurations), Environments (cloud container templates), Sessions (running instances with persistence), and Events (streaming interface). The architecture decouples the brain (Claude's reasoning) from the hands (tool execution in isolated containers).

Key features include persistent sessions, multi-agent orchestration, outcome evaluation ('dreaming'), persistent memory (97% fewer first-pass errors reported by early adopters), a credential vault for API keys, and Agent Skills compatibility.

Pricing and Comparison to Open-Source

Standard API token pricing plus $0.08/hour active session fees. A continuous agent costs roughly $57.60/month plus token costs. Compare to self-hosting an open-source harness on a $10/month VPS. Managed Agents eliminates infrastructure overhead entirely, but open-source harnesses like Hermes Agent and OpenClaw offer more flexibility at lower marginal cost for teams with infrastructure expertise.