# Build a Research Agent with Claude Managed Agents (Part 2 of 4) — Anthropic

## Executive summary

This workshop demonstrates building a sophisticated 'research desk' application using Claude Managed Agents (CMA). The system evolves from a simple console agent to a complex, multi-agent ecosystem capable of performing long-running, asynchronous tasks. Key architectural components include specialized skills (e.g., calling SEC EDGAR filings), subagent coordination, shared memory stores, and a self-supervision mechanism called 'Outcomes,' which uses a rubric and grader to ensure task completion and accuracy. The platform supports enterprise deployment through secure methods like Vaults and MCP tunnels.

## Key takeaways

- Multi-Agent Orchestration and Skills: Agents can be augmented with specialized skills (e.g., SEC EDGAR) and delegate tasks to specialized subagents. The main agent acts as an orchestrator, managing the workflow and coordinating multiple subagents to achieve a complex goal.
- Outcomes for Self-Supervision: Outcomes define a task, a rubric, and a grader. This mechanism enables asynchronous, long-running tasks by having a subagent evaluate the main agent's work against the defined rubric, allowing Claude to automatically retry until the outcome is satisfied.
- Secure Credential Management (Vaults): For production use, credentials are managed via 'Vaults.' Instead of injecting real secrets, a placeholder is used, which is swapped out for the real value only at egress on specified host domains, enhancing security.
- Shared, Long-Running Memory: The system supports long-running memory stores that can be shared across multiple sessions and agents, allowing the application to continuously learn and improve over time (a process referred to as 'dreaming').

## Technical details

- Agent Infrastructure and Deployment: The system can be run on Anthropic's infrastructure, AWS (via cloud platform), or self-hosted sandboxes (e.g., using providers like Burcell or Modal). The communication layer for chat-like applications is built over an event stream (SSE).
- Custom Harness Development: Developers can build custom orchestration harnesses on top of CMA to manage multi-session interactions, allowing the main agent to delegate work for multiple tickers or tasks simultaneously (map-reduce style).
- Security and Connectivity: Secure data access is achieved through Managed Cloud Platform (MCP) tunnels, which allow connecting private networks without exposing them to the public internet. Credentials are managed via Vaults, which inject placeholders that are swapped out at the point of egress.
- Agent Lifecycle Management: The process involves creating an agent and an environment (the sandbox template). The agent can be upgraded by adding skills and custom tools, which are treated as live artifacts that can be continuously updated.

## Practical implications

- Build engineers can design robust, scalable AI applications that move beyond simple chat interfaces by incorporating complex, multi-step workflows.
- The use of Outcomes provides a structured, auditable way to ensure AI tasks meet specific business requirements (rubrics) and can self-correct.
- Understanding the separation between the Agent SDK (a binary) and the managed cloud harness is crucial for choosing the right deployment model.
- The ability to define custom tools and skills allows for deep integration with proprietary or external APIs (e.g., SEC EDGAR).

## Topics

Build Engineering, AI Agents, Large Language Models (LLMs), Asynchronous Computing, Cloud Architecture, Data Security, Claude Managed Agents docs, Scaling Managed Agents: Decoupling the brain from the hands

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