--- title: "Claude Managed Agents: Anthropic Just Built the Agent Loop You Were Going to Write Anyway" date: 2026-04-10 tags: ["claude","managed-agents","agentic-workflows","anthropic","api"] categories: ["AI Tools","Agentic Workflows"] summary: "Anthropic launched Claude Managed Agents on April 8 — a managed API that handles the agent loop, sandboxing, checkpointing, and tool orchestration you'd otherwise build yourself. Here's what it actually offers, how the pricing model works, and why it matters for teams shipping production agents." --- Every team building a serious agentic workflow with Claude has written roughly the same thing: an agent loop. You manage the session state, wire up tool execution, handle errors and retries, implement checkpointing so a two-hour task doesn't restart from zero on a network blip, add tracing so you can figure out what went wrong in production, and bolt on sandboxing so Claude can't accidentally delete something important. It's a few thousand lines of plumbing that has nothing to do with your actual product. Anthropic shipped [Claude Managed Agents](https://claude.com/blog/claude-managed-agents) on April 8 in public beta. The pitch is simple: they built that plumbing for you. ## What It Is Claude Managed Agents is a suite of composable APIs for building and deploying cloud-hosted AI agents at scale. Instead of maintaining your own agent loop, you define what you want — the model, system prompt, tools, MCP servers, and guardrails — and Anthropic runs it on their infrastructure. The architecture is built around four core concepts: - **Agent** — the definition: model, system prompt, tools, MCP servers, skills - **Environment** — a configured cloud container with pre-installed packages (Python, Node, Go, etc.) and network access rules - **Session** — a running agent instance within an environment, executing a specific task - **Events** — the messages exchanged between your app and the agent: user turns, tool results, status updates You create an agent once and reference it by ID across as many sessions as you need. Sessions run asynchronously and stream responses back via server-sent events. You can send additional events mid-execution to steer the agent, or interrupt it entirely. ## Built-In Infrastructure You'd Otherwise Wire Yourself What makes this more than a thin wrapper around the Messages API is what Anthropic handles for you in every session: **Checkpointing.** Long-running sessions persist their state. A four-hour research task doesn't restart from scratch because a container hiccupped. The infrastructure handles recovery transparently. **Prompt caching and compaction.** The harness applies Claude's prompt caching automatically and manages context compaction for long sessions — two optimizations that significantly reduce cost and latency on extended tasks but require careful engineering when you implement them yourself. **Tool execution sandbox.** Claude can run bash commands, read and write files, search the web, and call MCP servers inside a secure container. You don't have to build the execution environment or worry about escape vectors. **Credential and permission management.** Scoped permissions and identity management are built in. You define what the agent can access; the runtime enforces it. **End-to-end tracing.** Every tool call, model turn, and event is logged and viewable in Claude Console. When something goes wrong in production, you have the full trace to debug against. **Multi-agent coordination** (research preview). Agents can delegate work to other agents. This is early — request access required — but it's the architecture needed for complex orchestrated workflows where a planning agent fans out to specialist subagents. ## The Pricing Model The cost structure is straightforward: standard Claude API token pricing plus **$0.08 per session-hour** of active runtime, measured in milliseconds. Idle time doesn't count. Web search, if your agent uses it, costs an extra $10 per 1,000 searches. For context: a complex agent task that runs for 20 minutes of active computation costs $0.027 in session runtime before token costs. For most production workloads, token usage will dominate the bill by a large margin — the session fee is noise unless you're running thousands of long sessions daily. All Managed Agents endpoints require the `managed-agents-2026-04-01` beta header. The SDK sets it automatically. ## Who's Already Shipping With It Anthropic announced several early adopters at launch: - **Notion** — custom agents embedded in their product - **Rakuten** — enterprise agents at scale - **Asana** — the AI Teammates feature that offloads project work to Claude agents - **Vibecode** — AI-native app development workflows - **Sentry** — automated debugging and patch generation The Sentry use case is worth noting specifically. Debugging production errors, generating reproduction cases, writing and testing patches, opening PRs — that's a multi-hour, multi-tool workflow that requires exactly the kind of stateful, long-running infrastructure Managed Agents provides. Building that on the raw Messages API is non-trivial. With Managed Agents it's a few hundred lines of configuration. Anthropic's internal testing showed up to 10 percentage points improvement in task success compared to standard prompting, with the largest gains on complex, multi-step problems. The gain isn't from the model — it's from the harness: better tool execution, automatic retries, and optimized context management that you'd have to implement yourself otherwise. ## What's Still in Research Preview Three significant features require separate access: **Outcomes / self-evaluation.** Claude iterates until it reaches defined success criteria rather than stopping after a single pass. This is the difference between "run until done" and "run until correct" — a meaningful distinction for quality-sensitive workflows. **Multi-agent coordination.** Agents delegating to other agents. The obvious use case is a planning agent that fans out work to specialist subagents and aggregates results — exactly the architecture Anthropic's own research on [agent teams](/posts/claude-code-agent-teams-multi-agent-orchestration/) pointed toward. **Memory.** Persistent memory across sessions. Currently, each session starts fresh; memory would allow agents to build up context across interactions over time. These are the features that push Managed Agents from "hosted agent loop" to "autonomous agent platform." The fact that they're research preview means they're real and being used, not vaporware — but they're not ready for general production use yet. ## Messages API vs. Managed Agents Anthropic's own comparison is honest about the tradeoff: | | Messages API | Managed Agents | |---|---|---| | **Control** | Full | Configuration-level | | **Infrastructure** | You build it | Anthropic provides it | | **Best for** | Custom loops, fine-grained control | Long-running tasks, async work | If you need to implement unusual agent architectures, deeply customize the loop, or have compliance requirements that preclude running agent state on Anthropic's infrastructure, the Messages API is still the right choice. For the majority of production agent use cases — tasks that run for minutes to hours, need tool access, and need to survive failures gracefully — Managed Agents removes the infrastructure work without constraining what you can build. ## Why This Matters Every serious team building production agents with Claude was going to reinvent this wheel. Some already have. Anthropic's Managed Agents offering isn't just a convenience layer — it's a statement about where agent infrastructure is heading. The model of "you write the loop, we provide the model" worked fine for inference. It starts to break down when the loop itself needs to be reliable at scale: stateful, recoverable, observable, and secure. That's real engineering, and it compounds with every new agent use case. By absorbing that complexity into the platform, Anthropic shortens the path from "Claude can do this" to "we ship this to customers" from months to days. For the ecosystem, that means faster iteration and more production deployments. For Anthropic, it means deeper integration into the workflows that matter. The research preview features — outcomes, multi-agent, memory — point at what comes next: agents that aren't just long-running but genuinely persistent, collaborative, and self-correcting. The plumbing is in place. The interesting question is what gets built on top of it. **Access:** Claude Managed Agents is available in public beta to all Claude API accounts. Request access to research preview features (outcomes, multi-agent, memory) via the [Claude platform](https://claude.com/form/claude-managed-agents). --- *Sources: [Claude Managed Agents launch post](https://claude.com/blog/claude-managed-agents) · [Official documentation](https://platform.claude.com/docs/en/managed-agents/overview) · [SiliconANGLE coverage](https://siliconangle.com/2026/04/08/anthropic-launches-claude-managed-agents-speed-ai-agent-development/) · [The New Stack](https://thenewstack.io/with-claude-managed-agents-anthropic-wants-to-run-your-ai-agents-for-you/) · [The Register](https://www.theregister.com/2026/04/09/anthropic_offers_to_host_ai/)*