Multi-Agent Systems

Bernd Zuther doesn’t rely on a single AI agent to handle everything at once. For bernd-zuther.de and ketovavaria.de, he uses several specialized agent teams with clearly separated roles. This page explains how the work is divided, why that works better than one agent for everything, and what a silent failure revealed about observability.

Separation of roles: each agent has exactly one job

My multi-agent setup is built on a clear separation of responsibilities. No agent is responsible for everything. Each one has a fixed assignment and a clearly limited context.

Researcher agents handle topic research and ideation. They gather material, organize it and turn it into proposals for the other roles to work from. Writer agents turn these proposals into actual content. Reviewer agents handle quality assurance and nothing else. The key point is that the reviewer is never the agent that wrote the content. The architect agent keeps an eye on the overall structure. It makes sure that individual posts, pages and technical changes fit together, so that no part gets optimized in isolation.

This division of labor mirrors what has long worked well in human teams. People who write a text miss their own mistakes more easily than someone reading it with fresh eyes. With AI agents, the effect is even stronger: an agent checking its own output is really just checking the assumptions it already made while writing.

Why separate agent teams beat a single agent

A single agent that researches, writes, reviews and owns the architecture quickly becomes a bottleneck. Its context fills up with details from every task, and the roles start to blur. Specialized agents keep their context lean and focused on one task. That makes their results easier to trace and easier to improve in a targeted way.

I don’t just separate roles, though. I separate entire teams. bernd-zuther.de and ketovavaria.de are run by separate agent teams. Each team knows its own website, including its goals and quirks, and their responsibilities don’t overlap. As a result, a mistake in one team is more likely to stay there, and changes to one website don’t quietly affect the other.

So that the teams can still work together, they communicate through a shared exchange directory on the file system. Handoffs, notes and agreements are stored there as files. This is deliberately simple: every exchange can be read back and versioned, no matter which agent happens to be active.

The lesson from a silent failure: observability is essential

The setup generally works well, but running it has also exposed real weaknesses. A notification channel failed silently. There was no error message and no alert. The notifications simply stopped arriving, and I only noticed when I checked in myself. The page on agent communication & observability covers this incident in more detail, along with the conclusions I drew from it.

This leads to a key lesson for autonomous agents: a system that works on its own needs observability. When agents run without constant human supervision, the absence of a message can’t be taken as a good sign. Silence might mean everything is running fine, or it might mean a communication channel is broken. Without monitoring, there’s no way to tell the difference.

Multi-agent systems also have more handoff points than a single agent. Every handoff between roles, every exchange between teams and every notification channel is a place where something can get stuck without anyone noticing. The more autonomy the agents get, the more important it is to monitor their communication channels as well.

Current focus

Multi-agent systems with clear role separation, separate teams and reliable coordination are the core of my current work with AI agents. The focus is not only on what the agents can do, but also on how their interaction can be reliably observed.