Claude Managed Agents Cheat Sheet

Claude Managed Agents is Anthropic’s hosting environment for AI agents.

Instead of constructing:

• orchestration
• sandboxes
• long-running sessions
• memory systems
• tool execution
• retries/checkpointing
• observability

…Anthropic manages the infrastructure for you.

Think of it as:

Core Concepts

1. Agent

An agent is the central AI worker that receives tasks, reasons through problems, decides which tools to use, and produces outputs.

An agent definition usually contains:

• the Claude model to use
• a system prompt or instructions
• allowed tools and permissions
• execution policies and constraints

The agent acts like an autonomous runtime process rather than a simple chatbot session.

Important characteristics:

• Agents can execute multi-step workflows autonomously.
• Agents can invoke tools repeatedly during reasoning.
• Agents maintain context throughout long-running sessions.
• Agents can collaborate with subagents for specialized tasks.

2. Environment

An environment is the secure execution workspace where the agent performs actions.

The environment can contain:

• filesystems
• installed dependencies
• API credentials
• secrets
• MCP servers
• runtime configurations

Claude Managed Agents uses isolated hosted sandboxes to reduce security risks and separate workloads.

Important characteristics:

• Environments isolate execution from local infrastructure.
• Credentials can be injected securely at runtime.

• Dependencies and runtime tools can be preconfigured.
• Multiple agents can use different environments depending on workload requirements.

Typical examples:

• A coding environment with Git and Python installed
• A research environment with browser access
• An enterprise environment connected to internal APIs

3. Session

A session represents the persistent execution state of an agent while it works on a task.

A session stores:

• conversation history
• intermediate reasoning context
• uploaded/generated files
• execution progress
• checkpoints and memory

Unlike normal stateless API calls, sessions allow agents to continue long-running workflows over time.

Important characteristics:

• Sessions preserve continuity across multiple steps.
• Sessions allow resumable execution after interruptions.
• Sessions enable long-context workflows without manually rebuilding state.
• Sessions support debugging and replay capabilities.

Example use cases:

• Multi-hour research workflows
• Repository-wide code analysis
• Long-running automation pipelines

4. Events

Events are structured runtime notifications emitted during agent execution. They provide visibility into what the agent is currently doing and what has already occurred during a workflow.

Events are important because Claude Managed Agents is fundamentally event-driven internally. As the agent reasons and executes tasks, the runtime continuously emits events describing state changes and execution activity.

Typical event types include:

• session started
• tool invocation started
• tool invocation completed
• message generated
• file created
• task completed
• error encountered
• checkpoint saved

Important characteristics:

• Events provide real-time observability into workflows.
• Events can be streamed while the agent is still running.
• Events help with debugging and execution tracing.
• Events enable external systems to react to agent activity. Events support asynchronous workflow architectures.

Common use cases:

• live progress dashboards
• audit logging
• workflow monitoring
• triggering downstream automation
• debugging failed executions

Why events matter:

Without events, developers would only see the final output after the workflow completes. Events expose the intermediate execution lifecycle, making the system more transparent and easier to integrate into production environments.

5. Tools

Tools allow the agent to interact with external systems and perform actions beyond text generation.

Examples of supported tools:

• shell/bash execution
• filesystem access
• web browsing
• API calls
• database queries
• MCP-connected applications

The agent uses tools iteratively during reasoning.

Important characteristics:

• Tool usage is model-driven and autonomous.
• Tools can return structured data for further reasoning.
• Agents may chain multiple tools together dynamically.
• Permissions can restrict which tools are available.

Examples:

• Running shell commands to inspect logs
• Reading source code files
• Querying GitHub pull requests
• Searching internal company documentation

6. MCP (Model Context Protocol)

MCP is Anthropic’s standardized protocol for connecting AI agents to external tools, applications, and data sources.

It functions similarly to a universal integration layer for agents.

Common MCP integrations:

• GitHub
• Slack
• Notion
• Figma
• PostgreSQL
• Google Drive
• internal enterprise systems

Important characteristics:

• MCP standardizes how tools expose capabilities to agents.
• MCP reduces the need for custom integration code.
• MCP supports secure and permission-aware tool access.
• MCP enables reusable integrations across multiple agent systems.

Why it matters:

Without MCP, developers often build custom wrappers for every service. MCP provides a shared interface that simplifies integration and interoperability.

Key Features

Hosted Sandboxes

Claude Managed Agents provides isolated hosted execution environments where agents can safely run commands and workflows.

Benefits include:

• improved security isolation
• reduced infrastructure management
• consistent runtime behavior
• simplified dependency handling

This is especially important for coding and automation agents that execute commands dynamically.

Persistent Sessions and Memory

Agents maintain execution continuity across long-running workflows.

Persistent memory enables:

• multi-step reasoning
• resumable tasks
• long-context analysis
• workflow checkpointing

This removes the need for developers to manually reconstruct context after every API call.

Autonomous Tool Orchestration

Agents can dynamically decide:

• which tools to use
• when to call them
• how to combine outputs

This enables complex workflows without requiring rigid orchestration logic from developers.

Built-In Observability

Claude Managed Agents includes monitoring and debugging capabilities for agent execution.

Observability features include:

• execution traces
• tool call history
• logs
• session inspection
• replay/debugging support

These features help developers understand how the agent reached decisions.

Checkpointing and Recovery

The runtime can preserve execution state and recover from interruptions.

This improves reliability for:

• long-running workflows
• unstable external APIs
• complex automation tasks

Instead of restarting entirely, agents can resume from saved checkpoints.

Subagent Support

Agents can delegate specialized tasks to other agents.

Benefits include:

• specialization
• modular workflows
• improved scalability
• clearer task separation

This architecture is useful for advanced autonomous systems.

Secure Credential Management

Environments can securely inject secrets and credentials during runtime.

Examples:

• GitHub tokens
• API keys
• database credentials
• cloud access tokens

This avoids hardcoding sensitive information into prompts or code.

Enterprise Integration Capabilities

Claude Managed Agents is designed to integrate with enterprise systems through MCP and APIs.

Common enterprise integrations:

• ticketing systems
• internal knowledge bases
• CI/CD pipelines
• source control platforms
• communication tools

This enables enterprise workflow automation without building custom orchestration layers from scratch.

Claude Managed Agents Pricing

Claude Managed Agents pricing is primarily usage-based. Costs depend on:

• model usage
• token consumption
• tool execution
• sandbox/runtime usage
• session duration
• additional platform services

Managed Agents pricing is separate from standard Claude API pricing because the platform also includes hosted execution infrastructure and orchestration capabilities.

Cloud Managed Agents Cheat Sheet Resources:

https://platform.claude.com/docs/en/managed-agents/quickstart
https://claude.com/blog/claude-managed-agents
https://platform.claude.com/docs/en/managed-agents/events-and-streaming
https://platform.claude.com/docs/en/about-claude/pricing