Show HN: Building a Deep Research Agent Using MCP-Agent (thealliance.ai)

Sarmad Qadri, creator of the open-source mcp-agent, shares his journey building a general-purpose deep research agent leveraging the MCP (Micro-Component Protocol) framework. His goal was to develop an agent capable of handling complex, multi-step workflows by dynamically calling a wide array of MCP-connected tools, enabling research tasks that span internal data warehouses and external data sources. The project evolved through three iterations—initially adopting a static Orchestrator pattern that planned all steps upfront, then an Adaptive Workflow with dynamic subagent allocation and external memory, and finally an improved Deep Orchestrator that blends planning with deterministic verification. The significance of this work lies in its pragmatic blend of LLM reasoning with classical software engineering principles such as deterministic plan validation and token-budget management, addressing common pitfalls like hallucinations, token inefficiency, and brittle planning. The Deep Orchestrator ensures plans are verified for dependency correctness and MCP server existence before execution, improving reliability over purely LLM-driven workflows. Additionally, Qadri’s focus on structured prompting—using functional, XML-tagged prompt composition—enhances clarity and maintainability in complex multi-agent interactions. This project underscores that simpler, verifiable architectures often outperform overly complex adaptive systems in deep research AI agents. Qadri’s modular, open-source approach offers valuable technical insights and reusable patterns for the AI/ML community seeking to build versatile agents that integrate LLMs with external tool ecosystems efficiently and robustly.