Claude Managed Agents Just Dropped – Here's What Changes

Key Concepts

• Managed Agents: Anthropic-hosted AI agents that manage their own infrastructure, environment, and session state.
• Environment: A sandboxed, persistent workspace where the agent executes code, installs packages, and accesses tools.
• Session: A stateful container that allows the agent to "remember" previous interactions and continue tasks across multiple turns.
• Tool Use: The ability for agents to execute bash commands, perform web searches, or use custom-defined functions.
• Praison AI: A framework that simplifies the creation and management of agents into a few lines of code.
• Credential Vaults: Secure storage within the Claude console for managing third-party integrations like Slack and Notion.

---

1. Managed Agents vs. Open Claude

The primary distinction lies in infrastructure management.

• Open Claude: Requires the user to self-host, manage servers, and handle technical configurations.
• Managed Agents: Anthropic handles the hosting, allowing developers to focus entirely on task execution and agent logic rather than server maintenance.

2. Step-by-Step Implementation (Programmatic)

To integrate managed agents into an application, follow these steps:

1. Installation: Install the library via pip install anthropic and export your API key from platform.claude.com.
2. Agent Creation: Use client.beta_agents.create to define the model, system prompt, and tool set.
3. Environment Setup: Use environments.create to define the workspace (e.g., "unrestricted" network access).
4. Session Initialization: Use sessions.create with the agent_id and environment_id to start a persistent conversation.
5. Execution: Use sessions.events.stream to send tasks and receive real-time responses.

3. Advanced Features and Customization

• Tool Control: Developers can enable or disable specific tools (e.g., bash, web search) during agent creation.
• Custom Tools: By defining a JSON schema, users can create custom functions (e.g., get_weather). The agent automatically identifies when to invoke these tools based on user input.
• Package Management: Environments can be configured to include pre-installed libraries like pandas or numpy for data processing tasks.
• Interruption: The system supports user.interrupt to stop a tool call or a running process mid-flight.
• Usage Tracking: Developers can monitor input_tokens and output_tokens to track costs and performance.

4. Simplified Development with Praison AI

The video demonstrates that the complex multi-step process can be condensed using the Praison AI library:

• Import Agent and ManagedAgent.
• Set the backend to managed_agent.
• Use manage.update_agent and agent.start(stream=True) to handle creation, environment setup, and session management automatically behind the scenes.

5. User Interface (Claude Console) Workflow

For non-programmatic setup or prototyping:

1. Templates: Select a template (e.g., "Support Agent") from the console.
2. Configuration: Define the environment and network access levels.
3. Integrations: Connect third-party services (Slack/Notion) via the Credential Vault.
4. Tracing: The console provides a visual "trace" of the agent’s thought process, showing step-by-step execution (e.g., searching knowledge bases, executing code, and sending Slack messages).
5. Deployment: Once configured, the console provides the necessary agent_id and environment_id to transition from the UI to production code.

6. Notable Quotes

• "Previously, AI answered questions, but now AI does the work. Your agent has an identity, a name, a role, and a home."
• "Managed agents remove the headache of managing your own infrastructure... you spend time making your agent perform tasks, rather than focusing on building the infrastructure."

Synthesis

Claude Managed Agents represent a shift from passive AI chatbots to active, autonomous agents capable of executing code, managing environments, and interacting with external APIs. By offloading infrastructure to Anthropic, developers can scale from single-agent prototypes to complex, multi-agent systems efficiently. The combination of programmatic control via the Anthropic SDK and the simplified abstraction provided by frameworks like Praison AI makes these agents highly accessible for real-world automation tasks.