Claude Announces: Managed Agents enters public beta today.
01
Have You Been Here Before?
You built an Agent. It runs flawlessly on localhost.
Then you try to ship it to production... and the alerts start flooding in:
Infrastructure nightmare: Alert platforms blaring, cloud function timeouts, memory overflows -- Agent has not started real work yet, you have already become operations
Customer Service Agent crashes, 47 tickets backlogged.
Cloud function timeout: 900 seconds.
Memory overflow killed by watchdog, circuit breaker tripped, retries exhausted...
Your Agent code is perfect, but you are stuck doing DevOps firefighting.
This is not an isolated incident. This is what nearly everyone encounters when deploying Agents to production.
Today Anthropic says: We will handle this dirty work.
02
Production-Ready in Just a Few Lines
Tell Claude what you want to build:
What do you want to build? Enter a request to automatically generate Agent configuration and API call code
Build an Agent that evaluates acquisition targets: research the company, pull financial data, run competitive benchmarking analysis, draft investment memos.
The system generates the YAML configuration, curl commands, and Session creation code automatically.
All you do is tweak the system prompt.
03
What It Handles for You
Managed Agents checklist: Sandboxing, Error recovery, Auth, Memory, Checkpointing, Retry policies -- all checked
Sandboxing, error recovery, authentication, state persistence, event management, file storage, resumable execution, retry policies...
Any one of these used to take weeks to implement.
04
Then the Agent Runs Itself
Build investment thesis for BuyCo -- Active Session: Agent scanning data room, reading financial reports, executing web searches, fully autonomous
Once the Session starts, you can see exactly what it is doing:
Scanning the data room file structure, opening P&L statements (421 million USD revenue, 59 million USD EBITDA), reading balance sheets (124 million USD net debt), searching retail industry benchmarks, pulling competitor EV/EBITDA multiples...
The entire process runs without anyone watching.
This is the core of Managed Agents: A long-running autonomous session, fully persistent, reconnects without losing state.
05
System Architecture
Claude Managed Agents architecture: Harness at center, connecting four modules -- Tools/MCP, Session, Sandbox, Orchestration
The architecture centers on a Harness (orchestration layer), connecting four modules:
Tools plus MCP: Built-in Bash, file I/O, web search and scraping, plus any MCP server integration.
Session: Each Session is an Agent instance with full history persistence.
Sandbox: Cloud container with Python, Node.js, Go pre-installed, internet-connected.
Orchestration: Multi-Agent coordination layer supporting task decomposition and parallel dispatch.
06
Early Customer Metrics
Rakuten
Five departments (Product, Sales, Marketing, Finance, HR) all integrated. Agents take requests via Slack and Teams, delivering spreadsheets, PowerPoints, and applications.
Key metrics:
Time-to-market: 24 days reduced to 5 days, a 79 percent reduction.
Each dedicated Agent deployed in under one week.
Code modification accuracy: 99.9 percent.
ML Engineer Kenta Naruse had the Agent run independently for 7 hours in vLLM (12.5 million lines of code), with numerical precision matching reference implementations perfectly.
7 hours. 12.5 million lines of code. Perfect numerical accuracy.
Previously, who would have dared throw that task directly to AI?
Yusuke Kaji, General Manager of AI for Business at Rakuten:
"With Claude Managed Agents, our senior users have become like Galileo, able to cross disciplinary boundaries and contribute in multiple directions. Each dedicated Agent deploys within one week, running long-duration tasks across engineering, product, sales, marketing, and finance in sandboxes, generating applications, proposal PowerPoints, and spreadsheets. As Agent capabilities grow stronger, Managed Agents let us scale safely without building our own Agent infrastructure, letting us focus entirely on democratizing innovation within the company."
Vibecode
Vibecode lets users build and publish apps through conversation on their phones, no coding required.
Their CEO says: Previously users had to manually set up LLM sandboxes, manage lifecycles, configure tools... a process taking weeks or months.
Now:
Cost to develop an app: 50,000 USD reduced to 100 USD.
Time: months compressed to under 1 hour.
Infrastructure setup speed: at least 10 times faster.
The 50,000 to 100 mentioned in the headline comes from here.
Sentry
Sentry originally had Seer, a debugging Agent that could analyze root causes, but stopped there -- developers still had to fix manually.
Now they have added the second half: Seer analyzes the cause, Claude writes the patch and opens the PR. Developers receive a ready-to-review fix.
Indragie Karunaratne, Senior Director of AI/ML Engineering at Sentry:
"Telling developers where the bug is not enough -- they want you to fix it. Now customers can go from Seer's root cause analysis directly to Claude writing the fix and opening a PR. The entire integration went from zero to production in just weeks, without maintaining custom Agent infrastructure."
Notion
Notion demo video: How Notion built with Claude Managed Agents
Dozens of tasks running in parallel. Engineers use it to write code. Knowledge workers generate web pages and PowerPoints. Teams collaborate on outputs.
Eric Liu, Product Manager at Notion:
"We want Notion to be the best place for teams to work with Agents. Managed Agents' ability to handle long-duration sessions, manage memory, and continuously output high-quality results makes this possible. Users can now delegate open-ended complex tasks, from writing code to generating PowerPoints and spreadsheets, without leaving Notion."
Asana built AI Teammates using Managed Agents, working alongside humans in Asana projects to accept tasks and draft deliverables.
Atlassian integrated Agents into Jira in just weeks.
General Legal's Agents can instantly generate code to retrieve any undefined query from uploaded documents, cutting development time by 10 times.
Blockit's pre-meeting intelligence assistant: automatically researches every participant, integrating calendar, contacts, and CRM data. From idea to production in days.
07
Pricing and Access
Pricing: Standard Claude Platform token costs, plus 0.08 USD per active session-hour.
Access requires the beta header managed-agents-2026-04-01. Official SDKs handle this automatically.
Claude Console includes built-in Session tracking, analytics, and debugging tools -- every tool call, decision, and error is visible.
The latest Claude Code includes a built-in claude-api skill. Just say "start onboarding for managed agents in Claude API" to get started.
Three features are in research preview, requiring separate application:
outcomes: Define success criteria for Agent self-evaluation and iteration.
multiagent: Multi-Agent coordination and parallel task distribution.
memory: Cross-session persistent memory.
Also launched same day: ant CLI, a command-line client for Claude API, supporting native Claude Code integration and YAML-based API resource versioning.
08
How Is This Different From Building Your Own?
Comparison: Messages API versus Managed Agents.
Use Case: Messages API suits custom control and fine-tuning. Managed Agents suits long-running tasks and asynchronous workflows.
Infrastructure: Messages API requires self-built infrastructure. Managed Agents offers fully managed infrastructure.
Execution Duration: Messages API handles shorter tasks. Managed Agents handles minutes to hours.
Cost: Messages API charges tokens only. Managed Agents charges tokens plus 0.08 USD per session-hour.
Both coexist. Teams with existing Agent architectures do not necessarily need to migrate.
But internal testing shows: On structured file generation tasks, Managed Agents achieved up to 10 percentage points higher task success rates than standard prompt loops, with the biggest gains on the hardest problems.
Many Agent projects die between demo and production.
Most die on infrastructure.
This time, Anthropic is saying: We will handle this dirty work!
Related Links:
https://claude.com/blog/claude-managed-agents
https://platform.claude.com/docs/en/managed-agents/overview
https://www.youtube.com/watch?v=45hPRdfDEsI