Build Your Own Deep Research Agent + Technical Writer Multi-Agent System

A hands-on workshop, presented at AI Engineering Conference Europe, building a multi-agent AI system with two MCP servers: a Deep Research Agent and a LinkedIn Writing Workflow. Both connected to a harness like Claude Code or Cursor.

🎬 Full workshop available on YouTube ↓

📑 Slides here.

Whenever You're Ready, Here's How to Go Deeper

This workshop is a 2–4 hour taste. If you want to go from zero to shipping production-grade AI agents, check out our Agentic AI Engineering Course, built with Towards AI.

34 lessons. Three end-to-end portfolio projects. A certificate. And a Discord community with direct access to industry experts and us.

Rated 5/5 by 300+ students. The first 6 lessons are free:

Start here →

How to Use This Repo

Three ways to use this repo. Pick the mode that fits the time you have. Or work through all three in order, since each builds on the last:

1. Watch the workshop and see the patterns end-to-end. Watch in ~2 hr. Start with the 2-hour YouTube workshop and the slides above. You'll come away with a mental model of the full multi-agent system: tool-use agents, evaluator-optimizer loops, grounded search, structured LLM output, and MCP-server design.

2. Run the finished code. See it produce real artifacts. Run in ~30 min. Watch the system generate a research brief, draft a LinkedIn post through an evaluator-optimizer loop, and score itself with an LLM-as-judge. Follow the Getting Started and Running the Code sections to install the project and run the MCP servers, skills, and evaluation pipeline.

3. Implement it yourself with agentic coding. Build a 1:1 replica from scratch in ~2–4 hr. Open implement_yourself/, a stripped-down skeleton prepared with 25 pre-groomed tickets and a custom /implement Claude Code skill that orchestrates SWE and Tester agents in a loop, ticket by ticket, until the directory matches src/. See implement_yourself/README.md for the kickoff guide.

   No cheating, by design. implement_yourself/ is a self-contained project. Open your harness (Claude Code, Cursor, …) directly in that folder (not at the repo root) so its working directory is scoped to the skeleton. The agents can't see the reference implementation in ../src/, can't grep it, can't read its files. You get a real build, not a copy-paste.

What You'll Build Today

Deep Research Agent — An MCP server that runs deep research using Gemini with Google Search grounding and native YouTube video analysis:

user topic → [deep_research] × N → analyze_youtube_video (if URLs) → [deep_research gap-fill] → compile_research → research.md

LinkedIn Writing Workflow — An MCP server that generates LinkedIn posts with an evaluator-optimizer loop:

research.md + guideline → generate post → [review → edit] × N → post.md → generate image

Both servers expose tools, resources, and prompts via the Model Context Protocol, letting any MCP-compatible harness orchestrate the workflow.

Patterns and concepts you'll learn:

• Tool-use agents — letting the LLM decide which tools to call and when
• Evaluator-optimizer loop — generate, review, edit in cycles
• Grounded search — Gemini with Google Search grounding for factual research
• Structured LLM output — Pydantic schemas for type-safe model responses
• MCP server design — registering tools, resources, and prompts with FastMCP
• LLM-as-judge evaluation — automated quality scoring with Opik

Example: End-to-End Workflow

Here's a real run through the full pipeline — from a topic seed to a published-ready LinkedIn post with an AI-generated image.

Final output

Phil Tobaloo AI Engineer | I ship AI products and teach you about the process. We planned 12 AI agents and shipped 1. It worked better. Sounds crazy, right? But it's a common story. A client built an AI marketing chatbot. Their initial design had dozens of agents: orchestrator, validators, spam prevention. It failed. A single agent with tools won. Tasks were tightly coupled. One brain maintained context. Tools were still specialized. This is the core mistake. People jump to complex multi-agent setups too fast. Think AI system design as a spectrum: Workflows: You control steps. Single Agent + Tools: Model decides flow. Multi-Agent: Multiple decision-makers. ... A single agent works for most cases. But it has limits. Too many tools? You hit "context rot." Past ~10-20 tools, LLMs degrade at tool selection. They get overwhelmed. Information gets lost in the middle. So, when do you actually need multi-agent? ... The simplest system that reliably solves the problem is always the best system. Don't overengineer your AI agents. Build simple first. What's the most complex agent architecture you've simplified? Tell me below. Read the full post

# Research Topic: AI Agent Architecture — When Less Is More

## Key Questions
1. Why do single-agent architectures with smart tools outperform multi-agent systems?
2. What are the only legitimate reasons to adopt a multi-agent architecture?

## References
- Stop Overengineering: Workflows vs AI Agents Explained (YouTube)
- From 12 Agents to 1 (DecodingAI article)

# LinkedIn Post Guideline

## Topic
Why most AI teams should use 1 agent instead of 12.

## Angle
Open with the counterintuitive "12 agents → 1" hook. Introduce the complexity
spectrum. End with a clear mental model.

## Target Audience
AI engineers and technical leads building LLM-powered applications.

## Key Points
- A team planned 12 agents but shipped 1 — it worked better.
- The spectrum: workflows → single agent + tools → multi-agent. Stay left.
- "Context rot": past ~10-20 tools, LLMs degrade at tool selection.
- Only 4 valid reasons for multi-agent.

## Tone
Direct, opinionated, engineer-to-engineer. No fluff.

Example 2 — Harness Engineering (click to expand)

Harness engineering isn't just a new term for prompt engineering. It's where AI is heading.

Agents got useful enough for code and tools, but they weren't reliable. They'd repeat mistakes. The bottleneck shifted from code generation to consistent, reliable behavior in real systems.

Think of it this way: prompt engineering is what to ask. Context engineering is what to send the model. Harness engineering is how the whole thing operates. It's the environment around the model, beyond just tokens.

Car analogy: the model is the engine. Context is the fuel. The harness is the rest of the car: steering, brakes, lane boundaries. It prevents crashes.

A harness includes tools, permissions, state, tests, logs, retries, checkpoints, guardrails, and evals.

Stop hoping the model improves. Engineer its environment. The burden shifts to us, the builders, to prevent repeat mistakes.

I use self-reflection in my Claude Code setup. The agent learns what I liked, saving tokens and time.

Real companies are already doing this. Anthropic's long-running agents externalize memory into artifacts. OpenAI built a 1M-line product with zero manual code using structured docs and agent-to-agent reviews. Stripe agents merge 1K+ PRs weekly within isolated environments. LangChain moved a coding agent from outside the top 30 to top 5 on Terminal Bench 2.0 by changing only the harness. Same model, better system.

This isn't just for coding agents. This is the new way software gets built.

The programmer's job is shifting: less writing code, more designing habitats for agents to work without issues. Think machine-readable docs, evals, sandboxes, permission boundaries, and structural tests.

Reliability is the real work. Not just prompting.

LLMs are heading into systems, workflows, harnesses. Value comes from orchestration, constraints, feedback loops—not just a single prompt. The future isn't one genius model. It's models in well-engineered environments.

That's why harness engineering matters. It's what happens when you stop demoing intelligence and start shipping it.

Want to learn more? I explain it all in my latest video: https://youtu.be/zYerCzIexCg What's your biggest challenge building reliable agent systems right now?

Example 3 — Angine de Poitrine for AI Engineers (click to expand)

Forget your latest AI model. There's a new system breaking the internet: Angine de Poitrine.

This masked duo from Quebec, deploys a high-resolution audio architecture. It makes everything else sound low-res.

Khn's custom double-necked microtonal guitar features 2x resolution: 24 notes per octave, not 12. Fine-grained frequency modulation.

Klek backs him up on drums, driving rhythms that feel like O(n^2) time signatures. Pure algorithmic complexity.

Their sound: "Dada Pythagorean-Cubist mantra-rock." Eastern traditions meet Frank Zappa.

Their 27-minute inference run for KEXP in Feb 2026 hit 7M+ views and broke the internet. Sold-out shows in NYC, London, Rennes followed.

Their anonymity adds another layer. Polka-dot costumes and papier-mâché masks are anonymous inference endpoints.

They communicate in an invented language. This ensures pure signal: raw output, no artist biases. A decoupled identity art experiment.

Khn's loop pedals are recursive pipelines. They stack complex guitar and bass in real-time, building dense soundscapes.

Just dropped: their new album, Vol. II, on April 3, 2026.

AI music is common. Angine de Poitrine proves human artistry is the ultimate non-deterministic function. Raw, complex, and deeply human.

You need to hear this.

What's the most complex system you've encountered recently?

Browse more full examples (seed, research, post drafts, reviews, final post + image) in the examples/ directory.

Tech Stack

Getting Started

Assumes working Python knowledge and basic familiarity with LLMs.

Prerequisites

Installation

1. Clone and configure:

   git clone https://github.com/iusztinpaul/designing-real-world-ai-agents-workshop.git
   cd designing-real-world-ai-agents-workshop
   cp .env.example .env          # add your GOOGLE_API_KEY (+ optional OPIK_API_KEY)
   

2. Install dependencies:

   uv sync
   

   Note: If you don't have Python 3.12+, uv can install it for you: uv python install 3.12, then re-run uv sync.

3. Verify the setup:

   make test-end-to-end          # runs research + writing pipeline end-to-end
   

   If it completes without errors, you're good to go.

Running the Code

There are four ways to run the workflows:

MCP Servers (recommended)

Connect the servers to an MCP-compatible harness (Claude Code, Cursor) for interactive use. This is the primary way to use the workshop.

Setup: The .mcp.json file is pre-configured. Both servers start automatically when you open the project in Claude Code or Cursor.

Usage:

1. Open the project in Claude Code or Cursor
2. Invoke an MCP prompt (e.g., research_workflow) to get guided through the full workflow
3. Or call individual tools directly for fine-grained control

Manual server start (advanced):

make run-research-server    # stdio transport
make run-writing-server     # stdio transport

Skills

Pre-built slash commands that orchestrate the MCP tools with sensible defaults. All output goes to outputs/{topic-slug}/.

Example:

/research-and-write

The skill will ask you for a topic and guideline, then run the full pipeline end-to-end. Check examples/ to see what each step produces.

Streamlit UI

A standalone chat UI that orchestrates both MCP servers via FastMCP — no harness required. Drop in a topic (or upload a .md / .txt seed file) and watch the pipeline run end-to-end with live per-stage progress: search counters, sources collected, evaluator-optimizer loop, and image generation.

make run-ui

Outputs land in outputs/{topic-slug}/ (same layout as the skills).

make test-research-workflow    # Research on a sample topic → test_logic/research.md
make test-writing-workflow     # Generate post from research → test_logic/post.md
make test-end-to-end           # Both steps sequentially

make run-dataset-writing           # Research + write for all dataset posts (with images)
make run-dataset-writing-no-image  # Same, skip image generation (faster)

Evaluation (requires Opik)

The workshop includes an LLM-as-judge evaluation pipeline. Instead of manually reviewing each generated post, an LLM scores them against quality criteria (structure, tone, accuracy). Opik tracks these scores across runs so you can measure whether prompt or pipeline changes actually improve output quality.

make eval-dev               # LLM judge on dev split
make eval-test              # LLM judge on test split
make eval-online            # Generate + judge posts on the fly

Each command automatically uploads the dataset to Opik before running. To upload without evaluating (e.g., to browse in the Opik UI), use make upload-eval-dataset.

Project Structure

├── src/
│   ├── research/              # Deep Research Agent MCP server
│   │   ├── server.py          # FastMCP entry point
│   │   ├── config/            # Settings, constants, prompt templates
│   │   ├── models/            # Pydantic schemas for structured LLM output
│   │   ├── app/               # Business logic handlers
│   │   ├── tools/             # MCP tool implementations
│   │   ├── routers/           # MCP tool, resource, and prompt registration
│   │   └── utils/             # Gemini client, file I/O, Opik, markdown helpers
│   └── writing/               # LinkedIn Writer MCP server
│       ├── server.py          # FastMCP entry point
│       ├── profiles/          # Shipped markdown profiles (structure, terminology, character, branding)
│       ├── config/            # Settings, constants, prompt templates
│       ├── models/            # Pydantic schemas (Post, Review, Profiles)
│       ├── app/               # Business logic handlers
│       ├── evals/             # LLM judge metric, dataset upload, evaluation harness
│       ├── tools/             # MCP tool implementations
│       ├── routers/           # MCP tool, resource, and prompt registration
│       └── utils/             # Gemini client, Imagen, Opik helpers
├── datasets/                  # LinkedIn posts dataset with labels and splits
├── examples/                  # Full end-to-end output samples (seed → research → posts → image)
├── scripts/                   # Entrypoints and test scripts
├── .mcp.json                  # MCP server configuration for harnesses
├── Makefile                   # Command center
└── .env.example               # Environment variable template

Next Steps

Contributors

Louis-François Bouchard AI Engineer	Paul Iusztin AI Engineer	Samridhi Vaid AI Engineer

License

MIT License. See LICENSE for details.

Copyright (c) 2026 Paul Iusztin, Towards AI Inc