Anthropic published details on June 10 of Claude Managed Agents, a production platform that runs agent orchestration, sandboxing, credential storage, session persistence, and observability on Anthropic’s own infrastructure. The company is no longer selling only model API calls. It is now selling the harness around the model as a managed service.
The strategic shift is significant. Since Claude opened to developers in 2023, the API contract was simple: tokens in, tokens out. Builders assembled their own loops, their own sandboxes, and their own state stores. That architecture gave teams full control and forced Anthropic to compete only on the model. Managed Agents changes the terms of that deal.
According to the Anthropic blog post, the core architectural change is decoupling the reasoning process from code execution. In a standard agent container setup, the model’s harness and the sandbox where code runs live inside the same container. Credentials sit next to the generated code. The container must spin up before the model can think. If the container dies, the session dies. Managed Agents separates those layers: the harness calls Claude independently, a separate sandbox handles tool execution, and an append-only session log connects the two. Anthropic says this cut time-to-first-token by roughly 60 percent at the median and by over 90 percent in the slowest cases.
The session layer is where the product becomes most interesting for operators. Every model call, tool call, and result is logged as an event outside the running process. Sessions can pause, resume, and be fully reconstructed from that log. Anthropic calls one feature “Dreaming”: a scheduled background process that reads session logs, extracts patterns, and writes curated memories back to the agent so it improves between runs without the developer managing a separate memory store.
Anthropic is competing directly with the agent-infrastructure ecosystem it previously left to third parties. LangChain and LangSmith, the sandbox vendors, the observability tooling built around OpenTelemetry export: all of these now have a first-party Anthropic alternative. The release announcement names Notion, Rakuten, Sentry, Asana, and Atlassian as customers already running production workloads on the platform.
The vendor logic is familiar. Make the hard parts invisible, make the convenient path also the fast path, and capture the relationship above and below the model. Builders who adopt Managed Agents hand Anthropic orchestration, sandboxing, credential management, and observability. In exchange, they get a harness that co-evolves with the model without requiring the team to retune it. Anthropic’s post gives a concrete example of that dynamic: on Claude Sonnet 4.5, agents exhibited “context anxiety” near the end of their context window, so the harness included resets to compensate. On Claude Opus 4.5 the behavior was gone and those resets became pure overhead. A team maintaining its own harness carries that tuning burden. A team on Managed Agents does not.
The lock-in question is real. Anthropic offers self-hosted sandboxes for teams that want code execution to stay inside their own VPC, and MCP tunnels for reaching private network servers. That architectural flexibility limits the blast radius of the dependency. But the session log, the Dreaming memory system, and the credential vault are Anthropic-operated by default. A team that builds deeply on those primitives is not moving easily.
The alternative path, assembling a custom harness on the Agent SDK or a third-party framework, preserves portability and control over the execution loop. That matters most to teams whose agent behavior is a genuine product differentiator, where the loop logic itself is proprietary. For teams building agents as a means to an end (automating internal workflows, accelerating support, drafting code patches) the trade looks different: give up the loop, keep the domain logic, ship faster.
Teams currently evaluating agent infrastructure should run a concrete comparison before committing. The 60 percent latency claim is Anthropic’s own benchmark, measured on Anthropic’s infrastructure, and independent results have not been published. Run your own task distribution against both architectures before a production commitment.
Anthropic (claude.com/blog), 2026-06-10.
