Done-for-you build
Section 1What Managed Agents Actually Are (and What They're Not)
Managed agents are not a new model or a new type of AI. They are a hosting layer. You define what an agent does - its tasks, tools, and guardrails - and Anthropic runs it inside their cloud infrastructure so you never have to set up the environment yourself.
What a managed agent actually is
The core mechanic
What the environment is
What a session is
What managed agents are NOT
❌ Not a replacement for Claude Code
❌ Not an always-on agent
❌ Not a stateful system (yet)
Who this is actually built for
The right user
The wrong user
Section 2How to Build Your First Agent in 4 Steps
The full setup from zero to a running agent takes under 10 minutes. No terminal. No code. You describe a goal and the interface builds the configuration for you.
Step 1: Create the agent
Where to go
Refining before you create
Adding business context
Step 2: Create the environment
What to configure
One environment per agent
Step 3: Connect credentials
How auth works
The vault
Step 4: Start a session and test
Running a test
What to check in the session log
Section 3Connecting Tools and Managing Credentials
Every managed agent connects to external tools through credentials stored in a vault. Connect once, reuse across agents. Your whole team can share the same vault.
The two ways to authenticate
OAuth (most tools)
Access token (manual)
How to set up and share a vault
Creating a vault
Sharing across the team
Adding new credentials
What to watch out for with CLI builds
❌ Don't let CLI agents hard-code API keys
Section 4What You Actually Get Charged On
Idle environments cost nothing. You only pay when a session is active and running. Keep sessions short and purposeful and the cost stays low.
The two cost components
Session compute: $0.08 per hour
Token costs
What costs nothing
Idle environments
No subscription required
How to estimate your costs
After a test run
Example from source
Section 5When to Use Managed Agents vs When to Skip Them
Managed agents solve a specific problem for a specific person. Outside that profile, they create friction without payoff. Run through the two questions below before you build anything.
Use managed agents when...
✅ You've never built an agent before
✅ You want team-shared credentials without DevOps
✅ You want simple task delegation inside a tool like Notion
Skip managed agents when...
❌ You already build in Claude Code
❌ You need the agent to run on a schedule
❌ You need the agent to remember previous runs
Section 6Limitations to Know Before You Build
Two gaps block most production use cases right now. Know them before you invest time building. Neither has a native fix yet.
Gap 1: No scheduled triggers
What the problem is
The workaround
What this rules out
Gap 2: No persistent memory
What the problem is
The partial workaround
What this rules out
Section 7Building Managed Agents From Claude Code CLI
Building from the CLI gives you one thing the Quick Start interface cannot: full project context. The agent you create will have a sharper, more specific system prompt because Claude already knows your business, goals, and stack.
Why build from Claude Code instead of the console
The context advantage
Real example from source
How to do it
The basic command pattern
How to use it as an orchestration layer
The credential warning
❌ Don't let CLI builds hard-code keys
Section 8What's Coming That Actually Matters
Three features are in early access. None are live yet - you have to apply to get in. All three close real gaps in the current product.
Outcomes: agents that check their own work
What it is
Why it matters
Multi-agent orchestration: coordinator delegates to specialists
How it works
What this enables
Persistent memory: agents that remember across sessions
What it is
Why it closes the biggest current gap