AI Agents Workshop: Deep Research and Writing Workflows

What changed

The iusztinpaul/designing-real-world-ai-agents-workshop GitHub repository offers a comprehensive, hands-on workshop focused on building multi-agent AI systems. The workshop details the creation of two primary agents: a Deep Research Agent and a LinkedIn Writing Workflow. The Deep Research Agent leverages Gemini with Google Search grounding and native YouTube video analysis to compile research into a research.md file. Its process involves user-defined topics, iterative deep research, video analysis if URLs are provided, gap-filling research, and final compilation. The LinkedIn Writing Workflow takes the compiled research and a guideline to generate a LinkedIn post, employing an evaluator-optimizer loop for refinement. This workflow includes generating a draft post, followed by multiple rounds of review and editing by an LLM-as-judge, culminating in a final post.md file and an AI-generated image.

The repository provides multiple ways to engage with the material: watching the full YouTube workshop (approximately 2 hours) to grasp the end-to-end system architecture, running the provided code to see real artifacts generated (around 30 minutes), or implementing a 1:1 replica from scratch using a stripped-down skeleton in the implement_yourself/ directory (2-4 hours). The latter approach enforces a strict learning environment by scoping the working directory to prevent agents from accessing reference implementations, ensuring a genuine build experience.

Key patterns and concepts covered include tool-use agents, where LLMs decide which tools to call; evaluator-optimizer loops for iterative generation and refinement; grounded search for factual accuracy; structured LLM output using Pydantic schemas; MCP (Model Context Protocol) server design for orchestrating workflows; and LLM-as-judge evaluation using tools like Opik for automated quality scoring.

Why it matters for builders

This workshop directly addresses the growing need for practical, production-grade AI agent development. Builders can learn to move beyond simple prompt engineering to architect complex multi-agent systems that can perform sophisticated tasks like in-depth research and content generation with iterative refinement. The emphasis on clear patterns and a spectrum of AI system design—from simple workflows to multi-agent setups—provides a valuable mental model for choosing the right architecture for specific problems, avoiding over-engineering.

Practical impact

Builders can leverage this repository to understand and implement advanced AI agent architectures. They can start by watching the workshop or running the provided code to see the agents in action. For a deeper dive, the implement_yourself/ directory offers a guided, hands-on experience to build a similar system from scratch, reinforcing learning through practical application. The detailed breakdown of the research and writing workflows, including example inputs and outputs, provides concrete examples that can be adapted for custom AI solutions. The discussion on the AI system design spectrum and the limitations of complex multi-agent systems offers crucial insights for efficient AI development.

Caveats and source limits

The source material is a GitHub repository containing workshop materials, not a formal product release or research paper. While it provides code, slides, and video, it does not include independent benchmark results for the agents' performance or specific pricing information for any associated services. The workshop focuses on conceptual understanding and practical implementation using provided tools and frameworks, with the understanding that actual deployment and scaling would require further engineering. The repository's star and fork counts (436 stars, 121 forks) indicate community engagement but are not direct performance metrics.