The failure, in plain terms
A four-agent pipeline — Researcher, Summarizer, Validator, Writer — was tasked with producing a competitive analysis. Each agent was given access to web search. Each agent was instructed to validate before passing downstream. Each agent performed its job correctly, according to its individual prompt.
The final output was confidently, completely wrong.
No individual agent hallucinated in a detectable way. The Validator gave the output a passing score. The system produced a high-confidence false consensus.
§01 — What happened — the trace
The Researcher queried for competitor pricing and returned three results. One of those results was an outdated cached page — competitor X showing their 2024 pricing. The Researcher had no instruction to check publication dates. It returned the result as valid.
The Summarizer received the three results and condensed them. The outdated result was a majority of the data. The summary reflected it.
The Validator was prompted to check whether the summary was internally consistent. It was. Every claim in the summary matched a source in the Researcher's output. Validator passed.
The Writer produced a polished analysis citing competitor X's pricing at the 2024 figure — a number that was 40% below current market rate.
The client used this analysis in a pricing meeting. They discovered the error three days later.
§02 — The mechanism — why this failure mode is structural
This is not a hallucination. No agent invented a number. The failure is a structural property of multi-agent systems where each agent validates against its immediate upstream, not against ground truth.
The pattern has a name in distributed systems: cascade validation failure. Each layer checks the layer above it for consistency, not for correctness. A consistent error passes every consistency check.
The specific trigger here is temporal validity — the question of whether retrieved information is current. Single-agent systems with web access often handle this poorly. Multi-agent systems amplify the failure because the outdated data is laundered through multiple "validation" steps, each of which increases apparent confidence.
By the time the Writer received the information, it had been validated twice. The confidence signal was strong. The underlying data was stale.
§03 — The three controls
Control 1 — Source dating as a first-class retrieval constraint.
Every tool call that retrieves external information must return a publication date alongside the content. If no date is available, the result is flagged as unverified. The Researcher's system prompt must include: "Reject any source that does not include a publication date within the last [N] days for time-sensitive domains."
Control 2 — Validator scope must include ground truth, not just consistency.
A Validator that checks internal consistency only is not a Validator. It is a consistency checker. Add a second validation pass: "Does this claim contradict any information retrieved independently? Run a verification query for the three most specific factual claims before passing."
Control 3 — Confidence must decay downstream, not accumulate.
The default behavior of multi-agent pipelines is confidence accumulation — each layer treats upstream output as authoritative. Invert this. Prompt every downstream agent: "You are receiving synthesized output from an upstream agent. Treat it as a draft hypothesis, not as verified fact. Your job is to challenge it, not to elaborate on it."
§04 — The uncomfortable truth
Most multi-agent validation pipelines are consistency pipelines wearing a Validator label.
Consistency is necessary but not sufficient. A system that is internally consistent and factually wrong is the most dangerous failure mode in production agent engineering — because it looks like it's working.
Before your next multi-agent deployment, ask one question: if my Researcher retrieved a confident, specific, completely false piece of information, which agent in my pipeline would catch it — and how?
If the answer is "the Validator," verify what your Validator's system prompt actually checks.
—— End of failure report ——
◆ Consistency is not correctness. A consistent error passes every consistency check.
◆ Ground truth validation beats upstream validation in every pipeline.
◆ Confidence should decay downstream, not accumulate.
— ORBIRESEARCH