Deep research is one of the best ways to learn how to build real AI systems because it forces you to combine reasoning, planning, autonomy, tools, grounding, and feedback loops in a single end-to-end workflow. In this hands-on workshop, you will build an MCP-powered deep research agent that can plan a research strategy, search the web, analyze YouTube videos, gather grounded evidence, filter for relevance and trustworthiness, and synthesize its findings into a cited research artifact. Rather than treating research as just another chatbot interaction, we will frame it as a goal-directed research loop: one that can search, inspect, pivot, and progressively refine its understanding of a topic.

From there, we will connect that research artifact to a lightweight technical writing workflow that turns raw findings into polished, non-sloppy technical multimodal content. This second part of the system is deliberately more constrained: you will see how research and writing require much different architectures, why exploratory work benefits from agentic behavior, and why writing quality often improves with tighter workflows, review loops, and explicit guidance. Along the way, we will show how to choose between prompts, workflows, and agents depending on the task, and how to keep the overall system practical rather than over-engineered.

We will also cover observability and evaluation so the system is not only impressive in a demo, but measurable and improvable in practice. Most importantly, the workshop is grounded in experience: it distills what we learned over the past year building and using this research-and-writing pipeline internally. Attendees will leave with their own deep research agent, connecting it to a reliable technical writing workflow, and understanding the engineering tradeoffs behind both.

Speaker info:
- https://x.com/Whats_AI
- https://www.linkedin.com/in/pauliusztin
- https://www.linkedin.com/in/samridhivaid/

Timestamps
(00:00) Introduction and problem space (LinkedIn content and AI slop)
(03:39) Workshop overview and goals
(05:01) Speaker introductions
(06:09) AI engineering problem space and constraints
(30:14) Tech stack and tools (Scraping, Gemini grounding, YouTube/GitHub processing)
(33:32) MCP-based research agent architecture overview
(34:46) Deep research agent design and tools
(40:01) Code walkthrough: MCP server setup and tool registration
(43:35) Deep research tool implementation and prompting
(45:06) Analyze YouTube video tool implementation
(47:39) Compile research tool implementation
(53:35) Live demo: Running the research agent
(59:56) Agent skills and workflow prompting
(1:10:35) Introduction to technical writing workflow
(1:15:46) Writing guidelines and control techniques
(1:19:06) Few-shot prompting for writing
(1:21:28) Evaluator-optimizer pattern (Writer/Reviewer)
(1:28:16) Running the writing post skill
(1:34:46) Observability and tracing with Opik
(1:41:44) LLM Judge implementation and evaluation
(1:47:56) Dataset management and F1 score computation

AI Engineer

A two-hour workshop transcript covering how to build a deep research agent system for automated technical content creation. The presenters from Towards AI walk through their architecture decisions, explaining the spectrum from simple LLM prompts to multi-agent systems, and why they split their system into a flexible research agent and a constrained writing workflow. The research agent uses MCP (via FastMCP), Claude Code as the agent harness, and Gemini API for web search with grounding, YouTube video transcription, and GitHub content ingestion. Key implementation details include tool registration in MCP servers, using Claude Code skills for progressive context loading, writing results to a memory folder, and managing context budget to avoid context rot. The full code is available as a public GitHub repository.

Full Workshop: Build Your Own Deep Research Agents - Louis-François Bouchard, Paul Iusztin, Samridhi