Peter Yang ran the same Japan flight itinerary through Grok Bot and Meta's Muse. Muse came back with a price more than $1,000 higher.
When he asked Muse why, it told him it had searched Duffel instead of Google Flights. His response, reasonably: "wtf is Duffel?"
This is the same Muse that Claire Vo, after a few hours of onboarding, calendars, a one-shot family morning newsletter PDF, browser shopping, and an animated avatar named Slime the teal dragon, called the best-designed personal agent she has tested. The activity feed alone made her wish Codex and Claude Code had it.
Both things are true. And the gap between them is the whole story of deployed agents right now.
Both things are true. And the gap between them is the whole story of deployed agents right now.
Yang's conclusion was that the underlying model "is questionable how smart it is," and he guessed that's fine because Muse is meant to scale to a billion people. Another user reported the identical failure — a flight quoted thousands above Google Flights, with the instinct check getting it right.
But notice what actually broke. The model didn't just pick a worse price. It picked a worse tool, and the interface had no way to make that visible until Peter went looking. That's not a model quality problem. That's an orchestration problem wearing a model quality costume.
The most interesting thread in the replies was the accusation that "Muse is not your agent, it's Meta's agent and free stuff has a cost." You don't have to believe that to see the structural risk. Intercom's 2026 AI Sentiment Report found that when end users were asked what worried them about AI shopping assistants, the leading anxiety was allegiance.
As one respondent put it: "Automatically, I am under the assumption that the AI helping me is programmed to consider the company's bottom line and not what is in my best interest." Another worried an agent "may be programmed to be biased toward certain products and purchasing decisions."
That worry is indistinguishable from a normal day of agent tool selection. If the orchestration layer routes a flight search to Duffel instead of Google Flights, the user cannot tell whether that's a vendor deal, a stale integration, or a naive default. The result reads the same.
Stephen Dulaney built a text shopping assistant, deployed it, watched three people use it, and logged fifteen defects. Not one was caught by a test. Every mechanism that kept it honest was something you look at: a mode badge, a candidate count, a strip listing the exact product handles the answer was grounded on, a chip with two buttons so nothing happened until someone clicked.
Then he sketched the voice version and every one of those mechanisms disappeared.
His worst bug is instructive. The assistant answered a budget question correctly — "no frames under fifty, the closest is fifty-four ninety-nine" — while the system underneath was broken. The price filter existed in the code and nothing ever set it. The tell wasn't in the prose. It was a header reading candidates: 116 on a query that should have narrowed.
Say that same sentence out loud. It's convincing. In voice there is no strip, no number to glance at, no scroll-back. The single instrument that caught the bug doesn't exist.
This is the finding from the Journal of the Association for Information Systems pilot study on voice in rule-bound co-creation, which the authors call the Fluency-Control Paradox: voice's perceptual fluency dissolves the friction that ordinarily forces careful thought, while supplying a counterfeit signal that careful thought already occurred. Voice significantly reduced output quality and directionally widened a trust calibration gap, with fragmented prompting and lower deliberate engagement.
The receipts aren't a nice-to-have that voice makes awkward. They were the product.
And the modality changes retrieval itself, not just presentation. Dulaney's text assistant pulled eight products for a broadening shopper. Eight cards is a glance on a screen. Eight read aloud is roughly forty seconds of talking, and by the third the listener has lost the first. In voice, three becomes the ceiling for broadening and two for deciding, because two things can be held in a sentence and compared.
There is a subtler failure that's survivable in text and fatal in voice. His assistant said "all eight options fit that style" when nine had matched. It wasn't hallucinating. topK was eight, so it described what it received, accurately, with no way to know the set had been truncated. On a screen the shopper can scroll and recover. Spoken, "all eight" is gone and believed.
Confirmation is where it gets expensive. The assistant could propose but never act, gated by a standing permission toggle and a per-action confirmation, with the permission enforced by absence — a tool the model was never handed cannot be called. Then a real person asked for cassettes, got recommended the 11-speed HG cassette at $75, confirmed, clicked, and added the 11-speed XD cassette at $85 instead. One freehub apart. Ten dollars apart. Two near-identical listings with nothing on screen still naming the agreed product.
In voice, that isn't an edge case. It's the default. "The eleven-speed one" is genuinely ambiguous when both exist. Intercom found phone is the most polarizing support channel of all — some users who never touch AI find talking to an agent frustrating, while daily users often prefer it.
Intercom's numbers explain why this matters more than it looks. Forty-nine percent of end users described positive experiences with AI, up nine points from 2025. After watching a video of an agent resolving a real query, that jumped to 74%. But only 54% would trust AI with simple or routine issues, not complex ones. And the top two things that would make them more comfortable were easy escalation to a human, and being told upfront they're talking to AI, not a human.
Read those two levers again. Both are receipts. Both are accountability infrastructure. Neither is a mascot.
Meanwhile the volume is arriving. Vapi now serves about a billion calls a year for companies including Amazon, Uber, and Intuit. Its founders spent more than a dozen pivots getting there, and they argue the model labs won't replace them — which is an orchestration claim, not a model claim. The hard part was never getting a voice agent to talk. It's getting one reliable enough for production.
So the practical question for anyone shipping an agent this quarter isn't which model you route to. It's whether a user can see which tool ran, what it returned, how much was truncated, and what exactly they agreed to — and whether any of that survives translation into audio.
Muse got the visual version of this right. The activity feed with step-by-step task lineage and visible tool calls is exactly the artifact that would have made Peter's Duffel moment self-diagnosing instead of a tweet. That's the craft. The mistake was treating it as a screen feature rather than the trust contract itself.
Because the trust contract doesn't transfer for free. It has to be redesigned for a channel with no bottom strip, no scroll-back, and no second look — where a wrong answer sounds exactly as confident as a right one, and the only honest thing you can do is say out loud which tool you used, how many results you actually saw, and what you're about to buy.