<p>Anthropic has launched Claude Managed Agents, a public beta service that lets developers build and deploy AI agents directly on Anthropic’s cloud infrastructure. Rather than spending months assembling sandboxed execution environments, credential stores, checkpointing systems, and tracing pipelines, teams can define agents in plain language or YAML and hand the infrastructure problem to Anthropic.</p>
<h2>What’s included</h2>
<p>The platform ships as a set of composable APIs organized around five areas: agent definition, session management, environments, credential vaults, and observability. Out of the box you get:</p>
<ul>
<li>Sandboxed code execution</li>
<li>Long-running autonomous sessions with checkpointing</li>
<li>Credential management with scoped permissions</li>
<li>End-to-end tracing</li>
<li>Multi-agent coordination (currently in limited research preview, along with advanced memory tooling and self-evaluation loops)</li>
</ul>
<p>Internal testing showed up to a 10-point improvement in task success compared to standard prompting loops.</p>
<h2>How it works in practice</h2>
<p>The interface has five tabs: Quick Start, Agents, Sessions, Environments, and Vaults. You describe what you want an agent to do in plain language, and the platform handles the rest. Agents can be refined iteratively without redeployment. No coding is required for basic use cases — people have built job recruiters that search listings and draft cold emails, news digest bots that post daily Slack summaries, and lead-capture receptionists, all without writing code.</p>
<h2>Pricing</h2>
<p>Billing is consumption-based: standard Claude API token rates plus $0.08 per active session-hour and $10 per 1,000 web searches. This introduces a three-dimensional cost structure — tokens, session runtime, and per-tool charges — that doesn’t map cleanly onto existing FinOps dashboards built for VMs or serverless workloads.</p>
<p>The attribution problem is real: autonomous agents running overnight have no natural cost ceiling, and current monitoring tools weren’t designed to track this kind of workload. Prompt caching can cut costs by up to 90% on repeated context, and model selection is worth evaluating on cost-per-outcome rather than cost-per-token. These are new primitives that need new accounting practices.</p>
<h2>Early adopters</h2>
<p>Notion, Rakuten, Asana, Sentry, and Vibecode are among the early adopters.</p>
<h2>What this signals</h2>
<p>This is Anthropic moving beyond model provision into the infrastructure layer for enterprise agent deployment. The bet is that the hardest part of shipping agents isn’t the model — it’s everything around it. By owning that layer, Anthropic positions itself closer to where production value actually gets created.</p>


Collections

Anthropic has launched Claude Managed Agents in public beta, a cloud platform that lets developers build and deploy AI agents without assembling their own infrastructure. The service provides sandboxed code execution, long-running sessions with checkpointing, credential vaults, multi-agent coordination, and end-to-end tracing via composable APIs. Agents can be defined in plain language or YAML. Pricing is consumption-based: standard API token rates plus $0.08 per active session-hour and $10 per 1,000 web searches, introducing a new cost model that existing FinOps tools aren't designed to handle. Early adopters include Notion, Rakuten, Asana, and Sentry. The move signals Anthropic expanding from model provider to infrastructure layer for enterprise agent deployment.

Claude Managed Agents launches in public beta