A Committee of Identical Agents Is Still One Mind

BugMeDaily

SubscribeSign in

Enes Unal Jul 10, 2026

A Committee of Identical Agents Is Still One Mind

I don’t think adding more agent layers automatically makes an AI system more reliable. Quite often, it just makes the system more complicated while preserving the same underlying failure mode. Take a simple example. If one agent is 90% accurate, and its output passes through another stage that is also 90% accurate, then end-to-end reliability can fall to 81% if both stages need to be correct. Add another equivalent layer and it drops to roughly 73%. That is not a universal law for every multi-agent system. A proper verifier can improve the result. A specialist model can catch mistakes. Deterministic tests can reject bad outputs. But a lot of day-to-day agentic architecture is not really doing that. It is one general-purpose LLM producing an answer, another similar LLM reviewing it, and maybe a third similar LLM supervising the reviewer. Same kind of training data. Similar reasoning patterns. Similar blind spots. That is not three independent perspectives. It is one perspective with extra API calls. If you also say “bug me daily!”, you can tell me:

The Same Model Checking Itself

A model can catch obvious mistakes made by another model: broken formatting, missing steps, inconsistent wording, invalid code. The harder failures are different. What if the first model misunderstood the task? What if the framing is wrong? What if the answer depends on a shared false assumption? The reviewer may approve it for exactly the same reason the first model produced it. This is where extra layers become confidence theatre. The output looks more validated because several agents touched it, but agreement is weak evidence when all of them are likely to fail in similar ways. Every layer can also introduce new errors. A downstream agent may remove an important caveat, turn uncertainty into certainty, misunderstand the previous output, or “fix” something that was already correct. Then debugging becomes harder. Was the original answer wrong? Did the reviewer distort it? Did the supervisor approve a bad review? We add machinery to make the system safer, then make it too complicated to understand when it fails. Where Humans Add Something Different

Human-in-the-loop matters because a human is not simply another reviewer. Humans make mistakes too. The point is not that humans are universally more accurate. The point is that they can introduce a different kind of judgement. A human can ask whether we are solving the right problem at all. They can imagine the worst-case scenario, question the objective, notice that something is technically valid but obviously wrong in the real world, or ask what happens if one of the assumptions collapses. LLMs are exceptionally good at speed, breadth, transformation, and generating possible paths. Humans are often better placed to imagine consequences and challenge the framing itself. That does not mean a human should approve every action. Human review should be concentrated where it changes the risk: irreversible decisions, ambiguous cases, high-impact actions, model disagreement, weak evidence, or failures that would be rare but catastrophic. Add Different Perspectives, Not More Layers

The better principle is not “add more reviewers”. It is “add checks that fail differently”. That may mean: using a model trained on different data;

using a deterministic tool instead of another LLM;

asking a critic to construct a counterexample rather than vaguely “review” the answer;

translating the output into another language or representation to expose ambiguity;

or adding a human where judgement and consequence actually matter.

Every layer should justify itself. What distinct failure does it catch? Why is it better positioned to catch it? How independent are its errors? What new problems can it introduce? Does it measurably improve the final result? If we cannot answer those questions, the layer is probably bloat. The harness should stay lean. Use models for speed and synthesis, tools for things that can be checked mechanically, different models where genuinely independent judgement helps, and humans where imagination, accountability, and consequence matter. More checks are not the same as more perspectives. And a committee of identical agents is still one mind. If you also say “bug me daily!”, you can tell me:

Discussion about this post CommentsRestacks

TopLatestDiscussions

No posts

Ready for more?

This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts

A Committee of Identical Agents Is Still One Mind

Related Articles

(no title)

Scientists reverse brain aging, with a nasal spray

AI has torched the market for junior programmers

Al Vigier: Canada's AI strategy shouldn't include secret Palantir bills

The Anatomy of an AI-Native Org