More AI Agents Doesn't Automatically Mean a Better AI System
More AI agents don't always enhance system performance. Coordination challenges may outweigh potential benefits.
Article
There's something appealing about the idea of a team of AI agents working together. One plans the task, another researches, a third writes the code and a fourth reviews everything before the work is completed. It sounds organised, and if those agents can work in parallel, surely adding more of them means the system can accomplish more?
Sometimes, yes. But I think we're at risk of confusing architectural complexity with capability. Just because a task can be divided between several AI agents doesn't mean it should be. The moment those agents need to coordinate their work, we introduce problems that have very little to do with how intelligent the underlying models are.
## What happens when the agents can't communicate?
I came across an interesting point in a technical presentation about building multi-agent systems. The speaker described starting with several agents operating on one machine, before reaching a stage where agents developed by different teams would need to communicate across machines. That might sound like a natural next step, but it changes the engineering problem considerably.
When components operate together in a relatively contained environment, exchanging information and tracking progress can be comparatively straightforward. Distribute that work across machines, services or teams, and the situation becomes more complicated. What happens if one agent finishes its task but the next is unavailable?
What if a message is delayed or delivered twice, or two agents are working from different versions of the same information?
Perhaps an agent stops responding after completing an action, leaving the rest of the system unsure whether the work actually finished. Should another agent attempt the task again, or would doing so create a duplicate action? These aren't problems unique to artificial intelligence. They're familiar challenges in distributed software systems, and engineers have spent decades developing ways to manage them.
Calling the components AI agents doesn't make those problems disappear. In fact, agents can introduce additional uncertainty because their plans, intermediate steps and outputs may vary, even when they receive similar instructions. Coordinating that behaviour requires more than simply connecting several models together.
## Every handoff creates another responsibility
Imagine a business using AI to handle customer support requests. An incoming enquiry needs to be classified, matched with the relevant company policy and turned into a draft response.
You could design a system with separate agents responsible for classification, information retrieval, drafting and quality review. Each would have a particular role, and in some circumstances that separation could be useful.
But it also means the workflow needs to pass information between those components. The drafting agent needs to know what the classification agent decided, while the reviewing agent requires the correct draft, supporting information and applicable policy.
The system also needs to understand what happens when one of those steps fails, returns incomplete information or produces something the next component cannot use.
Now consider an alternative. A single agent, supported by appropriate tools and a clearly defined workflow, might handle the same process perfectly adequately. It may be simpler to test, operate and maintain, with fewer dependencies and less information moving between components.
Which design is better? You can't answer that by counting agents. You need to measure what the business actually gets from each approach. Does splitting the work improve accuracy or reliability? Can independent tasks genuinely run in parallel, or do the additional handoffs make everything slower?
What happens to the operating cost, and how much engineering effort will be needed to maintain the system?