Anthropic’s managed agents
Anthropic announced Claude Managed Agents for production use with early integrations cited including Notion, Asana and Sentry and a listed price point around $0.08 per session‑hour ( ). The company frames these agents as tools for enterprise automation and says they’re tuned for cost and reliability as buyers shift budgets toward predictable deployment ( ).
Anthropic has opened Claude Managed Agents to developers as a hosted way to run long, tool-using artificial intelligence workers in production. (anthropic.com) The product lives in the Claude Platform and asks developers to define three pieces: an agent, an environment, and a session. Anthropic’s docs say the service is behind a `managed-agents-2026-04-01` beta header and supports pre-built tools including bash, file operations, and web search. (platform.claude.com) Anthropic describes the system as a split between “the brain” and “the hands”: Claude does the reasoning, while separate sandboxes run code and edit files. The company says it also separated session state, the orchestration loop, and the sandbox so each part can change without breaking the others. (anthropic.com) That design targets a common problem in agent software: a demo can work in a weekend, but production systems need durable state, tool routing, failure recovery, and isolated execution. Anthropic said it built Managed Agents as a hosted service for “long-horizon agent work,” meaning tasks that run across many steps instead of a single chat reply. (anthropic.com) Anthropic’s current pricing docs say Managed Agents sessions are billed for tokens on standard Claude model rates, adding a second billing dimension beyond model usage. Anthropic’s public materials tied to the launch also put the infrastructure charge at about $0.08 per session-hour, a pricing unit aimed at making long-running agents easier to budget. (platform.claude.com; x.com) The early customer list points to office software and developer tooling, not consumer chatbots. Anthropic has separately promoted work with Asana on “production-ready AI systems” and with Sentry on production deployment of Claude-based tools, while launch materials also named Notion among early integrations. (anthropic.com; anthropic.com; x.com) The release lands as Anthropic has been publishing a steady stream of engineering notes on how to make agents run longer and fail less often. In November 2025, the company wrote that agents asked to handle work spanning “hours, or even days” needed better harnesses, and in January 2026 it argued that agent evaluation gets harder as systems gain more autonomy and flexibility. (anthropic.com; anthropic.com) Anthropic’s pitch is narrower than “build any workflow with artificial intelligence.” The official docs focus on reusable agent configurations, versioning, controlled environments, and persistent event history, which are the kinds of controls large companies usually ask for before they connect a model to internal systems. (platform.claude.com; platform.claude.com) The next test is whether companies buy the managed layer instead of stitching together their own agent stacks. Anthropic is betting that predictable sessions, isolated sandboxes, and a posted hourly rate will sell better than another do-it-yourself framework. (anthropic.com; platform.claude.com; x.com)
