# Inside Managed Deep Agents: mda CLI, Context Hub, and Harbor Evals | Interrupt London

## Executive summary

Managed Deep Agents addresses the complexity of moving sophisticated AI agents from local prototypes to reliable, multi-tenant production systems. The platform tightly couples the agent's core logic (the 'harness') with robust infrastructure components (like LangSmith Deployment), allowing development teams to focus purely on business logic. Key features include centralized authentication, durable execution with checkpointing, version-controlled context management via Context Hub, and streamlined deployment using the `mda CLI`.

## Key takeaways

- Agent Evolution: From ReAct to Model + Harness: An agent is defined not just by its model, but by the 'harness'—the surrounding system that provides skills, memory, and system prompts. This harness is crucial for providing the model with the correct context at the right time.
- Managed Deep Agents Simplifies Production Deployment: The platform abstracts away complex infrastructure concerns like multi-tenancy, durable execution, and task queuing, which are typically difficult to manage when building agents from scratch.
- End-to-End Agent Lifecycle Management: The process is streamlined via the `mda CLI`: `mda init` scaffolds the project, `mda dev` runs it locally in LangSmith Studio, and `mda deploy` handles production deployment.
- Advanced Governance and Context Control: The system supports granular control over identity and context, allowing developers to define whether an agent uses user-scoped or agent-scoped credentials, and to manage personal vs. shared memory based on the calling channel (e.g., Slack DM vs. shared channel).

## Technical details

- Agent Harness Components: A robust agent harness must manage four key areas: 1) Execution Environment (with first-class support for file systems and sandboxing); 2) Context Management (progressive disclosure, summarizing, and offloading large tool outputs); 3) Delegation (decomposing tasks to specialized subagents); and 4) Steering (human-in-the-loop approval).
- Authentication and Identity: Managed Deep Agents use a centralized auth handler service to establish the principal for a call. Developers can explicitly define whether the agent should use user-scoped or agent-scoped credentials when interacting with tools or external services.
- Durable Execution and State Management: Leveraging the LangGraph runtime, the system supports durable execution and checkpointing, enabling agents to pause a run and wait indefinitely for user authorization (e.g., in Slack) and resume exactly where they left off.
- Context and Evaluation: Context is managed through LangSmith Context Hub, providing version control and a shared space for business users to collaborate on skills. Evaluation is simplified via first-class integration with Harbor, which can be initialized using `mda evals init` to set up checks and datasets.
- New Capabilities: Recent additions include HTTP channels (allowing custom channel integration), agent-created schedules (allowing agents to autonomously schedule reports or tasks), and per-run configuration for creating dynamic agents.

## Practical implications

- Build engineers can significantly reduce the overhead associated with building production-grade agent infrastructure (multi-tenancy, state persistence, task queuing).
- The standardized `mda CLI` provides a predictable, repeatable workflow for prototyping, testing, and deploying agents.
- Granular control over credentials and memory ensures agents can operate safely and compliantly across different organizational channels.
- The integration of Context Hub and Harbor evals allows for continuous, collaborative improvement and governance of agent capabilities.

## Topics

Agentic Workflow, Deep Learning Agents, LangChain, Build Engineering, Production Deployment, Context Management, CI/CD, LLM Governance, Managed Deep Agents, mda CLI, LangSmith Studio / LangSmith Deployment, LangSmith Context Hub, Harbor, Parallel Search

Source: https://www.youtube.com/watch?v=Wb3hQaG_xf0