Presentation: What I Learned Building Multi-Agent Systems From Scratch

2026-05-13 · Source: InfoQ · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

Paulo Arruda, a Staff Engineer at Shopify, details the company's progression in AI adoption, transitioning from basic chat tools to sophisticated multi-agent systems. He describes the shift from large, "all-in-one" prompts to specialized, narrow-focused agent microservices, which significantly reduced task completion times from hours to minutes. Arruda also introduces a forward-looking concept: using filesystem-based adapters, dubbed "llm-fuse," to manage context bloat in agents by allowing them to interact with diverse data sources as if they were local files. This approach aims to maximize precision and recall by injecting relevant knowledge and includes a "Defrag" tool for optimizing agent memory.

Key takeaway

For CTOs and VPs of Engineering grappling with scaling AI adoption and managing LLM context, your teams should prioritize developing lean, specialized agent microservices over large, monolithic prompts. This strategy, exemplified by Shopify's SwarmSDK, can dramatically cut task execution times and improve reliability. You should also explore adapter layers like "llm-fuse" to efficiently expose data to agents, mitigating context bloat and enhancing agent precision.

Key insights

Specialized, lean AI agents outperform monolithic prompts, drastically reducing task times and improving efficiency.

Principles

• Treat agents as narrow-focused experts, not generalists.
• Empower internal AI enthusiasts with tools, not a central "SWAT team."
• Minimize token usage for multi-agent orchestration.

Method

Transition from large, general prompts to multiple, specialized agent microservices. Implement a unified orchestration system (SwarmSDK) for in-process agent communication and workflow management. Utilize filesystem-based adapters to expose diverse data sources to agents as if they were local files.

In practice

• Break down complex tasks into smaller, agent-specific functions.
• Use multi-provider support for test generation and verification.
• Consider "llm-fuse" for efficient context management in agents.

Topics

• Multi-Agent Systems
• LLM Orchestration
• Shopify AI Strategy
• Agent Microservices
• Context Engineering

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Single Agent vs Multi-Agent: When to Build a Multi-Agent System — AI agents leverage LLMs, tools, and memory, often using the ReAct pattern, to perform complex tasks through reasoning and action.
• Mergeable by default: Building the context engine to save time and tokens — Peter Werry, Unblocked — Context engines optimize AI agent performance by providing highly relevant, curated organizational context, moving beyond simple data access.
• The AI scaffolding layer is collapsing. LlamaIndex's CEO explains what survives. — LLM application development is shifting from complex orchestration frameworks to context-centric, modular agentic systems.
• This Week in AI: Your Recap — Augmenting LLMs with advanced techniques like RAG and agentic workflows significantly enhances their real-world utility and precision.
• Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma, Qodo — LLM agents struggle with large contexts, planning loops, and task overload; structured multi-agent systems and context optimization are key.
• No AI Agent Orchestration Needed? What? — Modern frontier LLMs can self-orchestrate complex procedural tasks more effectively via in-context prompting than with external agent frameworks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by InfoQ.