Why More AI Agents Won't Fix Your System

There is a pattern playing out across AI teams right now that anyone who has managed a software project will recognize. A multi-agent system isn’t performing well, so the team adds more agents. A verifier here, a router there, maybe a second judge for good measure. Performance doesn’t improve. Sometimes it gets worse.

Nicole Königstein, CEO and Co-Chief AI Officer at Quantmate — an agentic quant research environment where trading strategies evolve from natural language prompts to live testing — has watched this pattern repeat across her own research and consulting work. She is simultaneously authoring two O’Reilly definitive guides on transformers and AI agents, and her PhD research focuses on self-improving AI agents. When she sees teams reaching for more agents as a fix, she reaches for a different reference entirely.

The Mythical Man-Month, but for agents

The parallel Königstein draws is to Frederick Brooks’s observation from 1975: adding more people to a late software project makes it later. The communication overhead grows faster than the output.

“Adding more people to a project does not necessarily speed up the project because they said over half the communication and you need to find a way of working around,” Königstein explains. “If you have maybe a very good team already working together, giving them more freedom to actually collaborate better together — this helps you solve the problem faster than just adding friction in adding five more people.”

The same dynamic plays out in multi-agent systems. Each new agent introduces token consumption for its own reasoning, but also for coordination — passing context, validating outputs, managing shared state. A five-agent pipeline doesn’t just cost five times what one agent costs. The overhead from communication, verification, and memory management compounds.

When a single model beats a fleet

Königstein references a DeepMind paper that validated what her own experiments had shown: overcomplicating the topology doesn’t always help.

“Not always is it necessary to overly complicate and have that extreme topology,” she says. “Sometimes if you are — you can also break it down into just subtasks and still have probably different LLMs, but I wouldn’t call it as in a multi-agent system. You guide the language model and maybe have smaller models involved and working together. And that might get a better outcome.”

The key decision isn’t whether to use agents. It is whether the coordination overhead is worth the specialization benefit. For many tasks — even complex ones — two or three models running sequentially with specific task guidance outperform a distributed agent network with supervisors, routers, and verifiers.

The real question to ask first

Before adding agents, Königstein recommends a different starting point: map the roles, not the agents. When consulting for companies, she asks them to describe the skills and tasks involved in their workflow, then works backward to the architecture.

“You need to ask yourself — every additional layer you add, every additional call you have, consumes tokens. If you have another verifier, it costs. If you have another check, it costs.”

This framing shifts the conversation from “how many agents do I need?” to “what is the minimum architecture that reliably produces the output?” The answer is often simpler than teams expect. A well-prompted single model with structured output validation can outperform a multi-agent system where the coordination logic itself introduces failure points.

What topology actually means for your system

Königstein thinks about multi-agent architecture through topologies — centralized, decentralized, blackboard — each with different coordination costs. A centralized topology (one supervisor routing to workers) concentrates decision-making but creates a bottleneck. A blackboard topology (agents sharing a common workspace) distributes reasoning but balloons memory consumption.

“Not every agent involved in that system needs to have every information about what the other agent is doing,” she notes. The architecture decision is really a memory and communication decision — and getting it wrong is more expensive than using fewer agents in the first place.

FAQ

Why does adding more AI agents make a system worse?

Each additional agent introduces coordination overhead — passing context, verifying outputs, managing shared memory. Nicole Königstein compares it to The Mythical Man-Month: adding people to a late project makes it later because communication costs grow faster than productive output. The same applies to multi-agent systems where token consumption compounds with every coordination step.

When should I use a multi-agent system versus a single model?

Use multi-agent only when task specialization clearly outweighs coordination cost. Königstein references a DeepMind study showing that simpler architectures — two or three sequential models with specific guidance — often outperform complex distributed agent networks. Map the roles first and ask whether each agent earns its token spend.

What is agent topology and why does it matter?

Topology describes how agents communicate and share state — centralized (one supervisor routing tasks), decentralized (peer-to-peer), or blackboard (shared workspace). Each creates different memory and token costs. Choosing the wrong topology can make coordination overhead exceed the value agents produce, according to Königstein’s research.

How do I reduce costs in a multi-agent AI system?

Start by questioning whether every agent is necessary. Königstein recommends evaluating early with domain experts, monitoring token consumption per agent, and considering whether sequential model calls with structured output could replace full agent coordination. Every verifier, router, and judge adds cost that compounds across the pipeline.

What is the biggest mistake teams make with multi-agent systems?

Assuming more agents equals better results. Königstein sees teams add verifiers, routers, and specialized agents without measuring whether the coordination cost exceeds the quality improvement. The fix is often removing agents and investing in better prompting and output validation for fewer, more capable models.

How do I evaluate if my multi-agent system is working?

Bring in subject matter experts early for stress testing. Königstein uses RULER, an open-source framework for relative evaluation, where a generator produces multiple outputs and a judge scores them against specific criteria. This provides a floating score with reasoning — catching quality issues before they compound across the agent chain.

What does The Mythical Man-Month have to do with AI agents?

Frederick Brooks observed that adding engineers to a late software project makes it later due to communication overhead. Königstein applies the same principle to AI agents: adding more agents creates coordination friction — context passing, state management, output verification — that can slow down and degrade system performance rather than improve it.

How do I decide between centralized and decentralized agent architecture?

Centralized systems (one orchestrator delegating to workers) are simpler to manage but create bottlenecks. Decentralized and blackboard architectures distribute reasoning but consume more memory and tokens. Königstein advises choosing based on your specific reliability requirements and whether agents genuinely need to share state or can work independently.

Can a single AI model replace a multi-agent system?

In many cases, yes. Königstein’s research shows that a well-guided single model with structured output validation can match or exceed multi-agent systems for tasks that don’t require true specialization. The key test: does breaking the task across agents produce measurably better results than sequential processing with one or two models?