Anthropic's Claude Managed Agents launches in public beta, promising to cut AI agent deployment from months to days with managed cloud infrastructure and built-in orchestration.
Anthropic launched Claude Managed Agents on April 8, 2026, a suite of composable APIs for building and deploying cloud-hosted AI agents at scale. The product is available now in public beta on the Claude Platform.
The launch targets a bottleneck that has kept enterprises from shipping AI agents in production: the months of infrastructure work required before users see anything. Notion, Rakuten, Asana, Sentry, and Atlassian are already deploying agents in weeks instead of months.
Building production agents required engineering teams to build secure sandboxing, checkpointing, credential management, scoped permissions, and end-to-end tracing from scratch. Claude Managed Agents handles all of that.
The product includes three core capabilities:
Session tracing, integration analytics, and troubleshooting tools are built directly into the Claude Console. Developers can access Managed Agents via SDK in six languages (Python, TypeScript, Java, Go, Ruby, PHP) or via the ant CLI.
Pricing is consumption-based: standard Claude Platform token rates apply, plus $0.08 per session-hour for active runtime.
The technical architecture decouples what Anthropic's engineering team calls the "brain" (Claude and its harness) from the "hands" (sandboxes and tools) and the "session" (the event log).
In the original single-container design, all three lived together. If the container failed, the session was lost. Credentials shared space with Claude's generated code, creating a security gap: a prompt injection could reach environment tokens and spawn unrestricted sessions.
The new design makes the harness call the container as a tool (execute(name, input) → string). The session log lives outside both. If the container dies, the harness catches it as a tool-call error and reinitializes a fresh one. If the harness crashes, a new instance resumes from the last event via wake(sessionId).
The security boundary is structural: credentials never reach the sandbox. Git repositories are cloned using access tokens during initialization, and MCP OAuth tokens are held in an external vault. Claude's generated code never handles the tokens directly.
The architecture gains are measurable. Anthropic reports that p50 time-to-first-token dropped 60% and p95 dropped over 90%.
Enterprise teams report shipping in days rather than months.
"Claude Managed Agents dramatically accelerated our development of Asana AI Teammates — helping us ship advanced capabilities faster — and freeing us to focus on creating an enterprise-grade multiplayer user experience." — Amritansh Raghav, CTO, Asana
Rakuten deployed specialist agents across engineering, product, sales, marketing, and finance in under a week per agent. Those agents plug into Slack and Teams, letting employees assign tasks and receive deliverables like spreadsheets and slides.
Sentry paired its existing Seer debugging agent with a Claude-powered agent that writes the patch and opens the PR. According to Indragie Karunaratne, Senior Director of Engineering AI/ML at Sentry: "Customers can now go from Seer's root cause analysis straight to a Claude-powered agent that writes the fix and opens a PR."
Atlassian is building agents that let customers assign tasks directly from Jira.
Notion launched a private alpha of Custom Agents, allowing teams to delegate coding and document generation without leaving the workspace.
"Before Claude Managed Agents, users would have to manually run LLMs in sandboxes, manage their lifecycle, equip them with appropriate tools, and oversee their execution — a process that could take weeks or months to set up. Now, with a few lines of code, users can spin up that same infrastructure at least 10x quicker than before." — Ansh Nanda, Co-founder, Vibecode
Anthropic is not the first to offer managed agent infrastructure. AWS Bedrock Agents provides multi-agent orchestration using foundation models, including Claude, integrated into AWS. Google Vertex AI Agent Builder offers similar capabilities on Google Cloud. Microsoft's Responses API underpins the Azure AI Foundry agent service.
OpenAI updated its Responses API in March 2026 to add a full computer environment, persistent terminal shell, and agent skill support.
Claude Managed Agents positions against these by being purpose-built for Claude's architecture and designed to evolve with the model's growing capabilities. Rather than encoding assumptions about what Claude can't do, the system decouples interfaces so the harness can be swapped as the model improves.
One concrete example: Claude Sonnet 4.5 would wrap up tasks prematurely as its context limit approached, a behavior the team called "context anxiety." Engineers added context resets to the harness. On Claude Opus 4.5, the behavior was gone and the resets were dead weight. With Managed Agents, those kinds of harness updates no longer require teams to rewire their agents.
Anthropic raised $30 billion in February 2026 at a $380 billion valuation, the second-largest private tech funding round on record. The company reports $14 billion in annualized revenue as of early 2026.
Claude's task-completion time horizons are expanding, according to METR benchmarks. As agents run longer, infrastructure reliability, security, and session durability become critical.
Claude Managed Agents is available now in public beta. Multi-agent coordination is in research preview, with access by request. Teams can start with the quickstart docs or from the Claude Console.
Cursor 3 introduces the Agents Window, a unified workspace for running parallel AI coding agents locally and in the cloud.
Google launches Gemma 4 in four sizes under Apache 2.0, built on Gemini 3 architecture. The 31B model ranks #3 among all open models on Arena AI.
Conductor raised a $22M Series A from Spark and Matrix. The Mac app for parallel AI coding agents has grown 10x since January.
