Why Managed Agents Are the Next Generation Core of Agent Building

LangChain has launched Managed Deep Agents in public beta, enabling easy building, running, and deployment of agents. Agent development has evolved from early AI frameworks and applications, through mature AI frameworks, to the current agent stage where LLMs invoke tools in a loop.

Recently, agent harnesses such as Claude Code, Pi, and Deep Agents have emerged, establishing ways to configure tools and execution environments suitable for agent loops. Simultaneously, durable execution for large-scale operations, sandbox for running untrusted code, and the 'brain and hands' architecture that separates an agent's judgment from its execution environment have become common designs.

Standards for controlling agent behavior have also advanced.

• AGENTS.md: Provides basic guidelines
• MCP: Connects to external systems
• Skills: Gradually reveals necessary context

The combination of such infrastructure and standards created the managed agent experience. This approach involves running harnesses on managed infrastructure while developers focus on business logic, including context, tools, and guidelines.

Building agents requires three elements.

1. Business logic provided by developers
2. Agent harness
3. Infrastructure to run the harness in production

Managed agents abstract the complexity of harnesses and operational infrastructure, helping developers focus on the actual behavior and business logic of agents.