Every multi-agent framework pushes the same playbook: spin up five specialized agents, let them debate, compress dialogue into shared memory, and let an evaluator model decide. In production, this design collapses: fresh benchmark data on Qwen2.5 and DeepSeek shows compressed debate memory consistently leads final models to wrong verdicts.
Here is what breaks under production load:
- Semantic Relevance Traps (+31% Error): Selecting memory by embedding relevance admits decoy facts because decoys share semantic space with criteria. In classic hidden-profile tasks, relevance filtering loses 23 to 31 accuracy points compared to rarest-first heuristics.
-
The Social Proof Bias (+24% Recovery): Final evaluator models suffer from algorithmic herd behavior, blindly favoring claims repeated across multiple agents. Stripping attribution tags (
[Agent-1],[Agent-2]) raises final decision accuracy by 13 to 24 points. - Transcript Context Bloat (5x Token Burn for Lower Accuracy): Passing raw or summarized debate transcripts burns 4,000+ extra context tokens per decision while yielding lower benchmark accuracy than providing zero discussion notes at all.
The One Real Trade-Off:
Stripping agent attribution prevents majority bias and hallucination spirals, but it destroys deterministic audit trails inside the LLM prompt. If you anonymize lines in the prompt buffer, you must maintain a cryptographic sidecar mapping to trace hallucinated tokens back to offending worker containers.
The Question to You:
What is your team using to aggregate multi-agent consensus? Are you stripping agent IDs before synthesis, or letting social proof dictate decisions? Drop your stack below.
Bohdan Hordii // Systems & Sovereign Infrastructure Architect | bohdanhordii.dev
Top comments (0)