Single Agent vs Multi-Agent: How to Actually Decide

Designing Agentic Systems | Episode 5 A leaking pipe needs a plumber. It does not need a plumber, an electrician, an architect and a project manager, unless the leak has reached the wiring and started coming through the wall. That is the whole question in agentic AI right now. Everybody is drawing multi-agent architectures. Very few people are asking whether the problem is complex enough to need one. In this episode I take the same failed fund transfer we have been investigating since Episode 1, build it first as a single agent, then watch it strain as the responsibilities pile up, and then rebuild it as a coordinator with three specialist agents. Then I give you the five questions I would ask before splitting anything. In this video What an agent is actually made of, and why the tools were already in your organisation before the AI arrived A single agent handling a failed transfer end to end, from customer complaint to reversal recommendation Why "agent" means two different things depending on who is saying it, and how to tell which one they mean The point where one agent starts getting overwhelmed, and what that looks like in a real banking journey The same investigation rebuilt as a multi-agent architecture with a coordinator What more agents genuinely cost you: configuration, operations, compute, context movement, conflicting conclusions, and an auditability problem Five questions that tell you whether to split The takeaway Split the system because the responsibilities are genuinely different. Not because a diagram with more agents looks more intelligent. Start with one agent. A single agent has one reasoning trail. A multi-agent system has several, plus a conflict resolver sitting on top of them. That is a real cost, and you should only pay it when the work actually demands it. Next episode Episode 6, designing for non-determinism. Guardrails, validation, retries, and how you contain an agent that goes off script. We will also talk about prompt injection, and if you have ever dealt with SQL injection you already understand more about it than you think. About this series This is Episode 5 of nine in Designing Agentic Systems, a series building towards one thing: a working agentic AI system inside an enterprise. Not a demo. Something that survives production, an audit, and a regulator asking why it did what it did. Every episode adds one more piece. Everything You Know About System Design Still Applies to AI Agents The New Component: What an LLM Agent Actually Is Agent vs Workflow vs Chatbot: Choosing the Right Pattern Anatomy of an Agent: Reasoning, Memory, Tools, Orchestration Single Agent or Multi-Agent (you are here) Designing for Non-Determinism: Guardrails, Retries, Failure Modes State and Memory in Agentic Systems Agentic AI in Banking: Credit, Fraud, Operations The Cost and Scale Reality: Tokens, Latency, Cloud 3.0 About me I am Ehson Golandaz, an Enterprise Architect with 23 years in banking and financial services. I write, mentor and teach on enterprise architecture, cloud, banking technology and AI. This channel is where I explain how modern systems are actually built. LinkedIn: https://www.linkedin.com/in/ehson-golandaz/ #AgenticAI #MultiAgent #EnterpriseArchitecture #AIAgents #BankingTechnology #SystemDesign