Anthropic just shipped Claude Managed Agents into public beta - a set of composable APIs designed to handle the infrastructure overhead that makes building production AI agents painful.
The pitch: instead of hand-rolling your own sandboxing, credential management, state tracking, and error recovery, you define the agent logic and Anthropic's infrastructure handles the rest. In internal testing, Anthropic says the approach improved task success rates by 10 percentage points compared to standard prompting alone.
Pricing is $0.08 per session-hour for active compute, with idle time billed at zero. For bursty workloads - agents that spin up, complete a task, then shut down - that could be substantially cheaper than keeping always-on infrastructure running continuously.
What the Managed Layer Handles
The infrastructure covers five things developers typically have to build themselves:
- Sandboxing - isolated execution environments so agents can't interfere with each other or production systems
- State management - the agent tracks what it has done across a multi-step task without you building a persistence layer
- Credential handling - secure storage and injection of API keys and tokens agents need mid-task
- Orchestration - routing work across agent steps in the right order
- Error recovery - automatic retries and failure handling when steps go wrong
Multi-agent coordination - where multiple Claude instances work in parallel on a task - is available in research preview, not yet in the main beta.
Production Customers Already Running It
Four companies are in production today: Notion, Rakuten, Asana, and Sentry. That's not a beta list of early-stage startups - Rakuten and Sentry handle real enterprise-scale workloads, which gives the infrastructure claim more credibility than a typical announcement-day list.
The "10x faster" development claim Anthropic is making refers to shipping time, not inference speed. The argument is that building sandboxing, state management, and error handling yourself adds weeks of engineering work. Offloading those to a managed service compresses the timeline considerably.
For developers already building on Claude, this lowers the barrier to production deployment. The DIY alternative - wiring up your own orchestration layer, your own isolated execution environment, your own failure handling - is genuinely complex work that has nothing to do with your actual agent logic.
One open question: Anthropic has published the $0.08/session-hour runtime cost but hasn't released a complete pricing breakdown that stacks Claude API call costs on top of infrastructure fees. The true cost per task depends on how long agents run, how often they're idle, and which Claude model you're calling - numbers that vary significantly by use case and will require real-world testing to pin down.