Anthropic launched Claude Managed Agents in public beta on April 8, 2026 — its first infrastructure-as-a-service product for running autonomous AI agents at scale. The platform abstracts away the engineering work that has kept most enterprises from shipping agents in production: secure sandboxing, state management across disconnections, scoped permissions, tool orchestration, and scaling. Pricing is $0.08 per session-hour plus standard Claude API token costs. Notion, Asana, Rakuten, Sentry, and Allianz confirmed adoption at launch. Anthropic claims 10× reduction in time from prototype to production and a 10-point improvement in task success rates versus standard API prompting loops.
Direct Answer: What is Claude Managed Agents and how does it work? Claude Managed Agents, launched in public beta April 8, 2026, is Anthropic’s managed cloud infrastructure for building and running autonomous AI agents at scale. Instead of engineers building their own sandboxing, state management, permissions systems, and agent orchestration — which typically takes months — Anthropic manages all of that infrastructure. Developers define what the agent should do, what tools it can access, and what guardrails constrain it. Claude handles execution, tool calling, error recovery, and state persistence. Agents can run for hours without disconnection, coordinate with sub-agents in parallel, and iterate until they meet defined success criteria. Pricing is standard Claude API token rates plus $0.08 per session-hour for active runtime (and $10 per 1,000 web searches if used). Early adopters include Notion, Asana, Rakuten, Sentry, and Allianz.
Why Enterprise AI Agent Deployment Has Been So Hard
Before Claude Managed Agents, the standard path to deploying a production AI agent went like this:
- Pick a model and write the agent logic — the interesting creative work
- Build a secure execution environment — a sandboxed container that can run code without affecting production systems
- Design state management — how does the agent remember what it has done if the connection drops?
- Implement permissions and tool access — what can the agent actually touch, read, write, or call?
- Handle errors and recovery — what happens when a tool call fails, when the context window fills, when the agent gets stuck?
- Build monitoring and observability — how do you know what the agent is doing, whether it is working, and when to intervene?
- Scale it — make it run reliably for many concurrent agents, not just one
This infrastructure stack typically takes a software team three to six months to build correctly. It is not AI work — it is DevOps and security engineering work. Most companies do not have the resources to do it well, and even those that do spend significant engineering time on infrastructure rather than agent capability.
This is the gap Claude Managed Agents targets. Anthropic handles the infrastructure. Teams focus on what the agent does, not on how it runs.
What Claude Managed Agents Actually Does
The platform is built around four core capabilities, each addressing a specific engineering challenge:
1. Secure Sandboxing
Every agent session runs in an isolated environment. Code the agent executes cannot access the host system, production databases, or other agents’ data without explicit permission grants. This matters for enterprise adoption: security and compliance teams need guarantees that an AI agent cannot inadvertently (or maliciously, if manipulated) access systems beyond its defined scope.
Sandboxing is the foundational requirement for deploying agents in regulated industries — healthcare, financial services, legal, government — where a misconfigured agent accessing the wrong system is not just an error but a compliance incident.
2. Long-Running Sessions with Persistence
Standard LLM API calls are stateless — each call is independent. Building an agent that works across a multi-hour task requires external state management: the agent needs to remember what it has done, what it has found, and where it is in the task.
Claude Managed Agents handles this internally. Sessions can run for hours and persist through disconnections. If the session is interrupted, the agent picks up where it left off rather than starting from scratch. Anthropic’s data shows Claude Code’s longest autonomous sessions have grown from 25 minutes to 45 minutes over three months — with Managed Agents, that ceiling is designed to extend significantly further.
An Anthropic research note published alongside the launch describes running Claude across approximately 2,000 sessions to produce a C compiler capable of compiling itself — a multi-step scientific computing workflow that demonstrates what long-running agent persistence enables.
3. Scoped Permissions and Tool Access
Agents need tools: web search, code execution, file access, API calls, database queries. The challenge is controlling which tools an agent can access, under what conditions, and with what scope.
Claude Managed Agents provides a permission model where developers define tool access explicitly. An agent processing invoices can access the invoicing database but not the HR database. An agent doing security scanning can read code but not write to production. The scoped permissions are enforced by the platform infrastructure, not just by prompting — making them robust against prompt injection and context manipulation.
4. Multi-Agent Coordination (Research Preview)
Complex workflows benefit from parallelism — multiple agents working simultaneously on different parts of a task, with results aggregated by an orchestrating agent. Setting this up manually requires significant infrastructure: message queues, agent state sharing, result aggregation, and failure handling.
Claude Managed Agents includes multi-agent coordination in research preview. An orchestrating Claude agent can spin up sub-agents, delegate tasks, and aggregate results — without the developer having to build the coordination infrastructure.
Pricing: The $0.08 Per Hour Question
The pricing model is consumption-based: standard Claude API token rates (which vary by model) plus $0.08 per session-hour for active runtime.
What does $0.08 per hour mean in practice?
For an agent running for one hour processing a complex document analysis task, the session cost is $0.08. The dominant cost is token usage — if the agent processes hundreds of pages and makes dozens of tool calls, token costs will far exceed the session fee.
For a long-running workflow agent working autonomously for 8 hours, the session cost is $0.64 — trivial compared to the human labour it replaces.
Web search costs $10 per 1,000 searches when agents use Anthropic’s managed web search tool. For research agents conducting hundreds of searches per session, this becomes a meaningful cost factor.
Comparison to building your own: A DevOps engineer in a high-cost market costs approximately $150,000–$200,000 per year. Building and maintaining agent infrastructure occupies a meaningful fraction of an engineer’s time. At $0.08/hour for the infrastructure, even at high agent utilisation, Anthropic’s platform is substantially cheaper than the engineering time alternative — which is precisely the pitch.
Early Adopters: What They Are Actually Doing
Anthropic confirmed five production deployments at launch:
Notion — collaborative workspace delegation. The knowledge management platform is using Claude Managed Agents to let Claude act as a workspace delegate — autonomously organising notes, summarising meeting outputs, and preparing document structures. This is the “AI colleague” use case: Claude works on shared Notion workspaces on behalf of human team members.
Asana — workflow automation. The project management platform is using agents to monitor project status, flag blockers, update task assignments based on capacity and priority, and draft stakeholder update messages. This is workflow intelligence on top of project data.
Rakuten — enterprise Slack agents via Claude Cowork. Japan’s largest e-commerce platform is running Claude agents inside Slack via Claude Cowork, handling internal requests, surfacing relevant enterprise knowledge, and routing queries to appropriate teams.
Sentry — automated debugging in production. The software error tracking platform uses agents to autonomously investigate production errors: pulling stack traces, correlating with recent deploys, checking related issues, and drafting root-cause analysis summaries. This is one of the highest-value AI agent use cases — debugging is time-consuming skilled work, and agents that can autonomously complete the initial investigation significantly reduce incident response time.
Allianz — customised insurance sector agents. Germany’s largest insurer is deploying agents for insurance-specific workflows — the details are not public, but the category suggests claims processing automation, underwriting assistance, or policy document analysis.
The Outcomes-Based Self-Evaluation Feature (Research Preview)
The most significant capability in research preview — not yet generally available — is outcomes-based iteration.
In standard AI agent workflows, the developer defines success criteria externally and evaluates whether the agent met them. With outcomes-based self-evaluation, Claude defines success criteria from the task description and iterates autonomously until it meets them — self-evaluating each output and deciding whether to continue refining or to declare the task complete.
“Claude self-evaluates and iterates until it gets there” is how Anthropic describes it. In internal testing on structured file generation, this approach improved task success rates by 10 percentage points over standard prompting loops, with the largest gains on the hardest problems — exactly where the improvement matters most.
This moves AI agents from “complete the task as instructed” to “complete the task successfully” — a fundamentally different capability with significant implications for autonomous workflow deployment.
The Competitive Landscape
Claude Managed Agents enters a market where major competitors have existing products:
| Platform | Provider | Key differentiation |
|---|---|---|
| Assistants API | OpenAI | First to market, broad ecosystem |
| Copilot Studio | Microsoft | Deep Office 365 integration |
| Vertex AI Agent Builder | Tight Google Workspace + Search integration | |
| AgentForce | Salesforce | CRM-native agents for sales/service |
| Claude Managed Agents | Anthropic | Safety defaults, Claude Code integration |
Anthropic’s stated differentiator is safety-focused infrastructure by default: sandboxing is not an option, it is the default. Permission scoping is enforced by infrastructure, not just prompts. For regulated industries deploying agents in sensitive contexts — the Allianz use case is representative — Anthropic’s safety defaults are a genuine competitive advantage over platforms where security is configured by the deploying organisation.
The Sovereignty and Privacy Assessment
For enterprise buyers with data sovereignty requirements, Claude Managed Agents is a cloud-hosted service. Agent sessions run on Anthropic’s infrastructure, and tool call outputs (including potentially sensitive document content, database query results, and API responses) pass through Anthropic’s systems.
This matters for:
- EU enterprises under GDPR: Data processed by Claude Managed Agents falls under Anthropic’s privacy policy and data processing agreements. The GDPR compliance documentation is available but requires review for specific use cases.
- Healthcare and financial services: Highly regulated data (PHI, PII, financial records) should not be processed by cloud-hosted agents without explicit DPA review and appropriate data residency guarantees.
- Government deployments: Most government AI use cases require on-premise or sovereign cloud deployment — Claude Managed Agents, as a cloud service, does not satisfy this requirement.
For organisations with sovereignty requirements that cannot use cloud-hosted agent infrastructure, the alternative is self-hosted Claude via Anthropic’s API, combined with self-managed agent orchestration — which is precisely the engineering problem Claude Managed Agents is designed to solve, but which requires that engineering investment.
FAQ
How do I access Claude Managed Agents? Claude Managed Agents is available in public beta via the Claude Platform API. You need an Anthropic API key. The SDK automatically adds the necessary beta headers. You can configure agents via the Claude console, Claude Code, or CLI.
What is the pricing for Claude Managed Agents? Standard Claude API token rates (which vary by model — Claude Sonnet 4.6 is less expensive than Claude Opus 4.6) plus $0.08 per session-hour for active agent runtime. Web search tool calls cost $10 per 1,000 searches.
How is Claude Managed Agents different from just calling the Claude API with tools? The Claude API with tools requires you to build state management, session persistence, sandboxing, and multi-agent coordination yourself. Claude Managed Agents provides all of this as managed infrastructure. The agent can run autonomously for hours, persist through disconnections, and coordinate with sub-agents — without you building or maintaining any of that infrastructure.
Can Claude Managed Agents run for hours without supervision? Yes. Sessions are designed for long-running autonomous operation with persistence through disconnections. The agent’s progress is saved so it can resume without restarting from scratch. Anthropic’s longest Claude Code sessions have reached 45 minutes of autonomous operation; Managed Agents is designed to extend this to multi-hour workflows.
Is Claude Managed Agents suitable for handling sensitive healthcare or financial data? Potentially, but with careful review. Anthropic provides data processing agreements for enterprise customers. For healthcare (HIPAA) or financial (GLBA) regulated data, you should review the specific DPA terms and data residency guarantees. Government and national security use cases typically require on-premise deployment, which Claude Managed Agents does not currently support.
What is multi-agent coordination in Claude Managed Agents? Currently in research preview, multi-agent coordination allows an orchestrating Claude agent to spin up sub-agents, delegate parallel tasks, and aggregate their results. This enables complex workflows where multiple agent specialisations work simultaneously — for example, one agent researching, another drafting, another reviewing — without the developer managing the coordination infrastructure.
