# Build a Research Agent with Claude Managed Agents (Part 2 of 4) Priyanka Phatak & Gabe Cemaj, Anthropic | AI Engineer | 35:36 Source: https://www.youtube.com/watch?v=FGftGOzfohc Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/build-a-research-agent-with-claude-managed-agents-part-2-of-4 Published: 2026-10-12 Tags: agents, evals, memory, multi-agent ## TL;DR - Claude Managed Agents can host the agent, sandbox, event log, scaling, and session infrastructure while an application communicates through session events. - A research agent can use skills, subagents, custom tools, and shared memory to analyze SEC filings and produce structured reports. - Outcomes pair a task with a rubric and a grader, allowing the agent to retry when its work does not satisfy the requested result. ## Summary Priyanka Phatak gives a short console demonstration, then Gabe Cemaj builds a research desk with Claude Managed Agents. The application exposes a chat interface over the session event stream, where browser messages become events and server-sent events return agent activity to the UI. The agent is then expanded with an SEC EDGAR skill, specialized subagents, a self-reference, a custom dispatch tool, and a shared memory store. Gabe shows how an outcome defines a research task, a rubric, and deliverables. A grader reviews the completed work and makes the agent try again when the result fails. The Nvidia example uses subagents for financial extraction and risk analysis, writes a structured scorecard, and updates memory. The Q&A covers vaults that replace placeholder credentials only at approved domains, self-hosted sandboxes, Agent SDK differences, missing budget controls, session steering, memory portability, and MCP tunnels for private data. ## Key ideas ### Managed Agents removes infrastructure work from the application team [03:12](https://www.youtube.com/watch?v=FGftGOzfohc&t=192s) Priyanka explains that a created agent runs in the cloud, and the application only needs to start sessions, send events, and receive events. Anthropic manages the hosting, harness, sandbox management, scaling, and event-log storage. The console demo creates an agent from a natural-language request, configures a sandbox environment, installs packages, and starts a session that answers a cat-joke prompt. Production users can tighten access by limiting reachable domains or specific MCP servers. ### The workshop builds a research desk around an event stream [04:42](https://www.youtube.com/watch?v=FGftGOzfohc&t=282s) Gabe's application starts as a chat-like interface and then grows into a research system. The browser sends a user-message event when someone submits text. The backend listens to the session's server-sent event stream and forwards those events to the browser for display. This is enough to build a chat application, but the design also supports asynchronous work that can run for minutes, hours, days, weeks, or months. ### Skills and subagents divide financial research into specialized work [13:41](https://www.youtube.com/watch?v=FGftGOzfohc&t=821s) The research agent receives a skill for calling the SEC and analyzing filings. Gabe adds subagents with their own prompts, tools, and permissions, plus a self-reference that lets the main agent create another copy of itself. A self-referenced agent receives the prompt, tool descriptions, and schemas, rather than the parent session's transcript. The custom dispatch_analysis tool later lets the main agent start separate sessions for a larger orchestration workflow. ### Outcomes give an agent a task and a test for completion [17:36](https://www.youtube.com/watch?v=FGftGOzfohc&t=1056s) An outcome contains the task, a rubric, and the expected deliverables. After the main agent finishes, a subagent reviews the work against the rubric. If the result is accepted, the session completes. If it fails, Claude tries again. In the research desk, the outcome asks for analysis of a company's recent SEC filings, a memory update, and a structured scorecard saved at a specified location. ### The Nvidia example combines subagents, artifacts, and shared memory [19:00](https://www.youtube.com/watch?v=FGftGOzfohc&t=1140s) For the Nvidia analysis, the agent checks its memory before repeating research. It delegates work to a financial analysis extractor and a risk evaluator, then uses their results to create the requested JSON scorecard and update memory. The console shows each subagent thread, which gives the operator visibility into the work. A separate grading subagent checks the rubric, the session's work, and its artifacts before accepting the result. ### Skills can change after deployment and affect future sessions [22:40](https://www.youtube.com/watch?v=FGftGOzfohc&t=1360s) Gabe describes skills as live artifacts. In the demonstration, the SEC skill is updated to support another form. New sessions that reference the skill receive the new information. Previous sessions can also be reviewed to see how the agent performed. The workshop connects this process to dreaming, which currently analyzes transcripts and memories across sessions to improve the shared memory store. ### Vaults keep real credentials out of the agent prompt [23:53](https://www.youtube.com/watch?v=FGftGOzfohc&t=1433s) For the demo, SEC identity information is placed in the prompt, but Gabe describes vaults for production credentials. A vault stores a secret behind a placeholder and is attached to a session by its vault ID. At egress, the placeholder is replaced with the real value only when the request targets an allowed host. The same mechanism can bundle MCP or OAuth credentials, such as credentials for a Google Workspace CLI or Notion. ### Managed Agents has a hosted harness and leaves loop controls to the builder [26:08](https://www.youtube.com/watch?v=FGftGOzfohc&t=1568s) Gabe describes the Agent SDK as a headless, Claude Code-like binary that can run wherever the developer wants. Managed Agents is a hosted solution with its own more generic harness, inspired by Claude Code but designed beyond coding tasks. The product does not currently provide budget limits for tokens or spending, so custom multi-agent harnesses must handle loop controls themselves. Session steering is possible by sending more events, while rewind and session forking were described as future work. ### Memory and private data can remain portable or inside a customer network [29:01](https://www.youtube.com/watch?v=FGftGOzfohc&t=1741s) A memory store can be shared across sessions, and multiple stores can be attached to support user-level and organization-level knowledge. Gabe says memories can be extracted and moved to other systems, although other memory providers are not built in. Private company data can be exposed through an authenticated MCP or CLI, an MCP tunnel that avoids exposing the MCP to the internet, or a self-hosted sandbox connected to the company's network. ## Notable quotes - Gabe Cemaj: "Now what we've created is our self-supervision." (17:41) - Gabe Cemaj: "It doesn't copy the transcript of that session into the sub agent." (15:53) - Gabe Cemaj: "We don't currently have a way for like rewind and roll forward." (28:35) - Gabe Cemaj: "We don't currently support out of the box other memory providers in the session but you would just build an MCP or a custom tool to do so." (30:30) ## Tools & references mentioned - Claude Managed Agents - Anthropic - Claude Code - Agent SDK - SEC EDGAR - AWS - Vercel - Modal - Daytona - MCP - Google Workspace CLI - Notion - Nvidia ## Who should watch - You are building an agent application and want a hosted session, sandbox, event-log, and scaling model rather than assembling those services yourself. - Your workflow needs long-running research, multiple specialized agents, structured deliverables, and a way to check whether the result is complete. - You need to decide how credentials, shared memory, private network data, and custom orchestration should fit around a managed agent service. ## Related talks - [Scheduled Agents, Dreaming and Q&A: Claude Managed Agents (Part 3 of 4)](https://aietalks.com/talks/scheduled-agents-dreaming-and-q-a-claude-managed-agents-part-3-of-4) (Priyanka Phatak & Gabe Cemaj, Anthropic, 26:31) - [Build Your Own Deep Research Agents](https://aietalks.com/talks/build-your-own-deep-research-agents) (Louis-François Bouchard, Paul Iusztin & Samridhi Vaid, Towards AI, 1:57:03) - [Claude Managed Agents Q&A: Evals, CI/CD and Multi-Tenancy (Part 4 of 4)](https://aietalks.com/talks/claude-managed-agents-q-a-evals-ci-cd-and-multi-tenancy-part-4-of-4) (Priyanka Phatak & Gabe Cemaj, Anthropic, 18:58) - [Evolving Claude APIs for Agents](https://aietalks.com/talks/evolving-claude-apis-for-agents) (Katelyn Lesse, Anthropic, 13:25) - [Claude Agent SDK [Full Workshop]](https://aietalks.com/talks/claude-agent-sdk-full-workshop) (Thariq Shihipar, Anthropic, 1:52:25)