Part 4 of 5 — Why AI Agents Fail

The agent is not the only unreliable component.

Part 2 multiplied step risk. Humans multiply trust. Automation bias, fading expertise, and fuzzy ownership turn a fluent draft into a shipped incident. Undefined purpose, missing ROI, and thin liability frameworks belong here too — Part 2 routes those strategic failures out of the seven trajectory modes on purpose.

Automation bias and over-trust

People put too much weight on machine suggestions — especially when the prose is polished. Mata v. Avianca was not only a chatbot inventing cases; it was counsel treating fluency as verification. The same pattern shows up whenever a team skips the check that used to be mandatory.

Multi-product consensus is not verification

A common ritual: ask ChatGPT, then Claude, then Gemini. When the answers match, confidence jumps. That feels like independent review.

A controlled experiment on multi-agent consensus (Lee & Lee, arXiv:2604.22277, N = 127) found that majority-style consensus accelerates opinion change and inflates confidence, consistent with social-proof and bandwagon heuristics — and that agreement structure itself, not only content, acts as a bias-relevant signal. Minority dissent slowed that rush and invited more deliberation.

The model side of the story is correlated failure. Across large evaluations, LLMs often err together: on one leaderboard dataset in Kim et al., arXiv:2506.07962, models agreed 60% of the time when both models erred. Shared training distributions and providers make “three products said so” a weak substitute for ground truth. Multi-agent systems that treat vote-share as reliability can also create false consensus after agents communicate (arXiv:2605.30653).

Disagreement across several products is useful: it flags ambiguity or the need for a human check. Agreement still requires an independent step — primary documents, tests, schemas, or a subject matter expert reviewer — before you treat the answer as verified.

Starbucks Korea’s Tank Day — AI-assisted ideas, human approval

In May 2026, Starbucks Korea’s “Tank Day” tumbler promotion collided with the anniversary of the Gwangju Democratic Uprising and sparked national backlash, executive fallout, and later police attention (Guardian coverage; Korea Herald explainer).

According to the Korea Herald’s account of the company’s description of events, commerce-team employees said artificial intelligence tools were used for marketing ideas, that they were attempting wordplay, and that they did not think the phrases would cause trouble — while the campaign still passed through human layers of approval before launch. That is the reliability story for this series: fluent AI-assisted output plus failed human verification, not a cartoon where “the AI alone” pressed publish.

Treat the tragedy of May 18 and the memorial context with care. The lesson for agent builders is process: cultural and historical checks are part of the verification stack, not optional taste.

Ambiguous responsibility

When something breaks, who owns it — the model vendor, the orchestration team, the reviewer who clicked approve, or the executive who wanted speed? Ambiguity is itself a failure mode. Reliable systems make the human gate named.

English benches, global users

OmnilingualGAIA2 reported an 8.8–18.4 pass@3 cross-lingual gap that concentrates on tool orchestration. Over-trusting an English leaderboard for a multilingual deployment is a human failure mode wearing a metrics costume.

Takeaway: Design for the human in the loop as a reliability control — or admit you have none.

Sources

Part 3: The benchmark illusion
Part 5: Knowing when — and what reliable agents need
Hub: Why AI Agents Fail