Claude Managed Agents: how to build and deploy production AI agents without the infrastructure headache

What You'll Learn

• How to set up your first Claude agent via Console without writing a single line of code
• Why most teams stalled at the prototype stage (and what Managed Agents fixes)
• The pricing math for different session volumes
• Real executive workflows: weekly reports, lead routing, content moderation, customer support triage
• How multi-agent coordination works and when to use it

Most executive teams hit the same wall when they try to build AI agents. They get a prototype working in a demo, sometimes in an afternoon. Then they try to put it in production. Suddenly they need sandboxed execution, persistent memory across sessions, credential management, permission scoping, and a way to trace what the agent did when something goes wrong. None of that is in the API. All of it has to be built from scratch.

That infrastructure problem is why "AI agents" have been a productivity promise for two years without becoming a productivity reality for most businesses. The teams that shipped production agents either had dedicated engineering resources or spent months patching together tools that weren't designed to work together.

Anthropic launched Claude Managed Agents in public beta on April 8, 2026. It's a hosted platform that handles the infrastructure so teams can skip from workflow idea to live agent in days rather than months. Notion, Rakuten, and Asana confirmed it at launch: all three reported 10x improvements in time-to-ship. Here's exactly how to get started.

Step 1: Understand what Managed Agents actually handles

What the platform covers: sandboxed code execution (agents run in isolated environments, so a bad agent can't touch your other systems), long-running sessions (agents that work for hours without timing out or losing context), checkpointing (agents can resume from where they left off after a failure), credential management (agents can use third-party tools without you hardcoding API keys), scoped permissions (each agent only gets access to what it needs), multi-agent coordination (agents can spawn specialized subagents and parallelize work), and full execution tracing (you can see every step the agent took).

Before Managed Agents, each of these required a separate engineering solution. Now they're included in the platform.

Step 2: Get Claude Platform access

Go to console.anthropic.com. If your organization doesn't have a Claude Platform account, create one. The Managed Agents beta is available to all Claude Platform users (no separate application required). Once inside, go to the Agents section in the left navigation. All API calls to Managed Agents require the beta header: managed-agents-2026-04-01 .

The platform uses standard Claude token pricing plus $0.08 per active session-hour. For a team running 5 agents for 2 hours each on weekdays, that's roughly $4 per day in runtime costs on top of token usage.

Step 3: Define your first agent workflow (the most important step)

Before touching the Console, write out your workflow in plain text. Answer three questions: (1) What trigger starts the agent? (a schedule, an incoming webhook, a user action), (2) What data does the agent need to access? (file storage, a database, an external API), (3) What does the output look like and where does it go? (a Slack message, a document in Google Drive, a row in a spreadsheet).

For a first agent, pick the simplest possible version. A good first agent: every Monday at 8 am, pull last week's data from [source], summarize it into 5 bullets, send to the team Slack channel.

Step 4: Build via Claude Console (no-code path)

Open the Console, click New Agent. You'll see: a system prompt field (define the agent's role and behavior), a tools section (add web search, code execution, file access, or custom API connections), and a trigger configuration (schedule, webhook, or manual). Write the system prompt like you'd write a job description.

Example system prompt:

Then set your trigger, connect your data sources, and run a test session. Review the execution trace to confirm the agent did what you expected.

Step 5: Understand session management and cost control

Active session-hours are billed at $0.08/hour. An agent that runs for 15 minutes counts as 0.25 session-hours ($0.02 in runtime). Set session time limits inside the agent configuration: under "Advanced Settings," set a maximum session duration. For most reporting agents, 30 minutes is generous. For research agents processing large document sets, budget 2-4 hours.

Use scoped permissions to limit what each agent can access (an agent that reads reports doesn't need write access). Review execution traces weekly for the first month to catch inefficiencies before they accumulate in your bill.

Step 6: Move to multi-agent coordination for complex workflows

Once your first agent runs reliably, consider whether breaking it into subagents improves reliability or speed. Multi-agent coordination works best when: (1) different parts of the workflow need different tools or permissions, (2) some steps can run in parallel (saving time), (3) a specialist subagent produces better output than a generalist single agent.

Example: a competitive intelligence workflow with three subagents (one that searches the web and extracts mentions, one that analyzes sentiment and flags key developments, one that formats everything into a Slack-ready brief). The orchestrator agent coordinates all three and handles the final output.

Action Steps Summary

• Get Platform access. Visit console.anthropic.com, create or log into your Claude Platform account, and navigate to the Agents section. Managed Agents is available in public beta to all Platform users.
• Define workflow on paper first. Write down your trigger, required data sources, and expected output format before touching the Console. Clarity on paper saves iteration time later.
• Build via Console. Click New Agent, write your system prompt as a clear job description, connect your data sources, set your trigger (schedule or webhook), and run a test session.
• Set session time limits. Under Advanced Settings, cap session duration (30 minutes for reporting, 2-4 hours for research). This prevents runaway costs from loops or stuck processes.
• Review the first 5 execution traces. Check each trace after the agent runs to spot inefficiencies, unexpected tool calls, or permission issues before they compound.
• Plan the upgrade to multi-agent. Once your first agent stabilizes, evaluate whether a second subagent would improve speed or reliability. Start multi-agent only after you've debugged the single-agent version.