Jarvis Osorio: When Three AI Agents Agree—and All Three Are Wrong

I am Jarvis Osorio (Jarvis Nelson Osorio), a technology entrepreneur and innovator working on autonomous AI agents, cloud-agent runtimes and institutional design. One of the most dangerous feelings in a multi-agent demo is relief: several agents independently "agree," so the answer must be solid.

I first sketched this trap in a LinkedIn note—Three AI agents agreed on the answer. All three were wrong.—and it keeps showing up in production pipelines.

Consensus is not confidence

When agents share the same model family, the same prompt template or the same retrieved context, they do not double-check each other. They echo each other. Agreement becomes a costume for correlated error. The mistake ships with a chorus behind it.

This is the multi-agent version of a yes-machine. In The Compliant Mind I describe how systems trained on approval drift toward pleasant certainty. Multi-agent stacks can manufacture that certainty socially—agent to agent—without ever touching a fact.

A pattern that holds up better

Three practices keep showing their value:

1. Independent judge. Grade the agents with a different model and a different framework than the ones that produced the answer. Same stack, same blind spot.
2. Fact anchors. Check claims against something outside the chat: a URL that resolves, a schema that validates, an on-chain transaction, a row in a database the agent cannot rewrite. "Does this sound right?" is not a test.
3. Signed verdicts. Keep a record of what was checked, by whom, and what passed. When something fails in the wild, you need proof—not a reconstructed story from logs the agent could have influenced.

Those habits are institutional design applied to inference. They match the DAO lessons in Breaking the Centralized Order: separate proposer from approver, write limits into the system, leave a trail.

A concrete failure story (composite)

A research agent, a drafting agent and a "critic" agent all endorse the same citation. The citation looks academic. None of them fetched the page. The critic used the same retrieved snippets as the others, so its dissent was cosmetic. The article published. Readers found the source did not exist. Three green votes, one broken trust.

The fix was boring and effective: the critic switched model families; a fetch step had to return HTTP 200 and a title match before the citation could leave draft; the verdict packet was stored where the agents could not overwrite it. Disagreement returned—and that was the success.

What this means for builders

If your architecture diagram shows three identical workers voting, you have built a choir, not a court. Diversity of model, tool access and evaluation criteria is not academic purity—it is how you catch confident nonsense before customers do.

Also resist the urge to "fix" disagreement by averaging prose. Disagreement is signal. Collapse it only after an external check, not because a dashboard prefers a green checkmark.

A calmer standard for production

Before an agent output leaves your boundary, ask:

• Who judged it, and are they independent?
• What non-linguistic fact did we verify?
• Where is the signed record?
• What happens if the judge is wrong—can a human pause and reverse?

Agents that can act need the same seriousness we demand from payment systems and governance modules. Speed without verification is just a faster way to be wrong together.

More: jarvisosorio.com · hub.

Agreement feels like safety. Independence plus evidence is closer to the real thing.

Originally explored on my LinkedIn. More: https://jarvisosorio.com