Full Workshop: Build Your Own Deep Research Agents - Louis-François Bouchard, Paul Iusztin, Samridhi
TL;DR
This workshop demonstrates how to build production-grade deep research agents by navigating the spectrum from rigid workflows to autonomous systems, emphasizing strategic context management and hybrid architectures to automate high-quality technical content creation while avoiding generic 'AI slop'.
🚫 The AI Content Problem 3 insights
Avoiding Generic AI Slop
Standard LLM outputs suffer from hallucinations, outdated information, and overused phrases like 'intricacies' and 'rapidly evolving' that provide shallow, meaningless value.
Automation Solution
The team built a deep research and writing agent system to automate technical course and article creation, replacing the expensive manual process of research plus human writing.
Meta-Learning Approach
They used this same agentic system to generate a comprehensive course teaching students exactly how to build the deep research system itself.
🎛️ The Autonomy Spectrum 3 insights
The Autonomy Slider
AI engineering exists on a spectrum from simple prompting to complex agentic systems, where increased autonomy reduces control and raises costs per task.
Workflows Defined
Workflows execute predetermined steps in fixed sequences—like a ticket handler that always classifies, routes, drafts, and validate—offering reliability without dynamic environmental reaction.
Agents Defined
True agents autonomously plan actions, dynamically select tools, and react to environmental changes, making them necessary only when tasks require branching logic like a CRM marketing content generator.
🧠 Context Management & Architecture 3 insights
Context Rot Threshold
LLM performance degrades significantly around 200,000 tokens due to the 'lost in the middle' training problem, occurring well before theoretical context window limits.
Delegation Pattern
Manage context budget by delegating specialized tasks to tools or sub-agents with isolated system prompts and validation logic rather than overloading a single agent's memory.
Multi-Agent Triggers
Deploy multi-agent systems only when facing specific constraints like 20+ tools, strict security or compliance requirements, or insufficient single-agent context windows.
🔍 Deep Research System Design 2 insights
Core Capabilities
Effective deep research agents combine planning, autonomous web navigation, reliable source citation, and feedback loops to perform thorough research without human micromanagement.
Hybrid Implementation
Production-grade AI products integrate workflows, single agents, and multi-agent systems rather than using one pattern exclusively, with deep research serving as a prime example of this unified approach.
Bottom Line
Start with the simplest workflow possible and only increase autonomy when tasks explicitly require dynamic decision-making or branching, while aggressively managing context windows through delegation to prevent performance degradation.
More from AI Engineer
View all
Think You Can Build a Game with AI? Think Again! - Danielle An & David Hoe, Meta
Meta engineers Danielle An and David Hoe argue that while AI has democratized basic game creation, true differentiation requires human taste, cohesive aesthetics powered by key art anchoring, and innovative runtime LLMs that enable unscripted, dynamically personalized gameplay experiences previously impossible in traditional development.
Beyond the Harness: A Journey Towards Adaptative Engineering - Rajiv Chandegra, Annicha Labs
Rajiv Chandegra introduces 'adaptive engineering,' a paradigm shift from fixed AI harnesses (like Cursor or Claude Code) to dynamic, self-organizing systems that emerge during runtime, enabling AI to handle complex, real-world messes beyond deterministic software environments.
What if the harness mattered more than the model? - Aditya Bhargava, Etsy
Aditya Bhargava argues that sophisticated agent harnesses can compensate for weaker open-source models, enabling local AI to match proprietary performance while reducing vendor dependency.
Frontier results, on device - RL Nabors, Arize
Rachel Lee Neighbors introduces a framework for replacing expensive cloud-based frontier models with Small Language Models (SLMs) running on-device, demonstrating how a systematic 'prototype big, deploy small' approach using evaluation tools like Phoenix can cut inference costs to zero while maintaining 90% accuracy and enabling offline functionality.