Building a Research Chat App on LangChain Managed Deep Agents (With Human Approval Before Web Search)

Key Takeaways

• Agent behavior defined in files, version‑controlled via Git.
• Human‑in‑the‑loop approval required for web‑search tool.
• Backend supports managed, local, and deployment runtimes via strategy pattern.
• Streaming SSE delivers tokens and interrupt events without WebSockets.
• Subagents act as specialist services, limiting main prompt blast radius.

Pulse Analysis

AI engineering is shifting from single‑prompt demos to full‑stack systems that must handle tool use, state, and reliability. The research chat app exemplifies this transition by treating the agent as a version‑controlled artifact stored in an "agent/" directory, while the FastAPI backend abstracts the execution environment. Whether the runtime points at LangSmith’s Managed Deep Agents, a local open‑source interpreter, or a custom LangGraph deployment, the front‑end sees a uniform API. This strategy‑pattern approach decouples business logic from infrastructure, enabling teams to swap runtimes without code changes and to embed the agent definition directly into CI pipelines.

A standout feature is the human‑in‑the‑loop (HITL) interrupt configured via "tools.json". By flagging the web‑search tool as requiring approval, the system emits an interrupt event over Server‑Sent Events (SSE) and pauses execution until a user explicitly approves or rejects the action. This policy‑as‑data model replaces hard‑coded conditional logic, providing a clear audit trail and preventing unintended external calls that could incur costs or expose sensitive data. The use of SSE for token streaming and interrupt signaling avoids the complexity of WebSockets while delivering low‑latency, responsive UI updates.

From a business perspective, the architecture delivers safety, cost predictability, and deployment flexibility—key concerns for enterprises scaling AI agents. Subagents, such as the fact‑checker, act as specialized micro‑services, containing blast radius and simplifying prompt engineering. The modular design also supports rapid iteration: updating "AGENTS.md" and re‑running the provisioning script propagates changes to the managed cloud agent, enabling continuous delivery. As organizations look to embed AI assistants into workflows, this example provides a reproducible blueprint for building, governing, and scaling trustworthy agent‑driven applications.