Leadership1 distinct publisher2 min readUpdated
Brian Distelburger told Business Insider his personal agent stack triggered a Resy deactivation without booking a single table. Recovery ran on goodwill, not support tickets.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
The portable part of this is the identity. He gave the agent his own Resy login [5], so nothing on the platform's side separated a piece of software from the customer whose name was on the account, and the penalty landed where the credential pointed [7]. Any agent handed live access to a vendor account inherits that property by default.
Getting back in is where the numbers are worth writing down. Distelburger says he had heard secondhand that Resy restores accounts after a month, so he waited the month out, and the month passed with nothing [12]. The letter that finally worked came from Hannah, the coordinating agent in a stack he runs on a Mac Mini with subagents for sports and for restaurants [13][2][3][4]. It drew a reply in two days [15]. That is roughly 32 days of outage against 48 hours of detection, about sixteen times longer to reverse than to cause [17][18]. The only messages that produced a substantive answer were the ones he did not write himself [19].
Look at what the successful letter had to reach for: Resy's policies, the length of his customer history, and his standing as an American Express Business Platinum customer [14]. None of that is available to an analyst who wired an agent into a shared vendor account last quarter. The appeal path for agent misbehaviour is currently the same path used for any commercial dispute, and it pays out on tenure and spend.
The interim was procurement. With no Resy account, he bought the highest tier of Dorsia, a competing reservation service that requires a minimum commitment, costs real money, and carries thinner inventory [9]; before that, he and his wife booked on her account [10]. Those are the workarounds nobody puts in a plan and everybody reaches for in the actual month.
What the episode lacks is a rule. Distelburger's own read is that most sites and companies have not worked out how they want agents to engage with their services [16]. The boundary is real, it is enforced by pattern detection, and it is lifted at someone's discretion. A single reinstatement, won by a well-argued letter from a long-paying customer, is not something a workflow can be designed around.
So the scoping question for putting agents into a workflow is not capability. It is which accounts can be gone for a month, and whether anyone at the vendor will pick up while they are.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Business Insider published an as-told-to essay based on a conversation with Brian Distelburger, 47, cofounder of Windmill, an AI performance review startup, who lives in New York.
Distelburger said he first built an OpenClaw instance on a Mac Mini with a personal agent, found OpenClaw complex, and moved to Hermes, also housed on a Mac Mini, which he has used for the past four or five months.
His primary agent is named Hannah, after his executive assistant of about 15 years; it coordinates subagents and is the one he interfaces with on Telegram.
His subagents include one dedicated to sports and one he calls Concierge, dedicated to events, restaurants and reservations.
He set up an agent with access to Resy and gave it his account credentials, so it could check for reservations at 4 Charles when they were released.
Within about 48 hours he received notice that his Resy account had been deactivated.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single self-reported first-person account
Every factual element rests on one edited as-told-to essay from the person involved. No Resy notice, terms-of-service text, email thread, agent log, or third-party confirmation is shown, and the key causal step — that bot detection caused the ban — is explicitly hedged as the subject's assumption. The internal timeline is coherent and specific, which supports the narrative facts, but nothing is independently verifiable.
Concrete but individual-scale usage
There is genuine, dated, sustained usage to point at — four or five months on a Hermes stack, a working multi-agent hierarchy, a real platform enforcement action, and a paid substitution to top-tier Dorsia — but all of it is one person's household. No organizational deployment, user counts, or evidence that platforms are seeing agent traffic at scale appears in the supplied material.
Mildly overstated causality, otherwise proportionate
The framing that 'another agent got me reinstated' asserts more than the timeline can carry, since the winning email was also the first one sent after the month his informal sources said precedes reinstatement, and Resy never explains its decision. Otherwise the essay is unusually restrained: it reports zero bookings, admits the guardrail failure, and hedges the detection mechanism, so the overstatement is modest rather than structural.
Founder visibility in an unchallenged format
The narrator cofounds an AI startup and the piece positions him as an early, aggressive adopter of agent tooling, which rewards a story where his agents are both the problem and the solution. The as-told-to structure gives Resy, Dorsia, and American Express no chance to contest the account. There is no evidence of a commercial relationship with any named platform or vendor, so the incentive is reputational rather than transactional.
Low-moderate
The narrative facts are specific, internally consistent, and self-incriminating in ways that argue against fabrication, so confidence in what happened to this one account is fair. Confidence in the interpretation — that agent traffic triggered the ban and that agent authorship won the appeal — is low, resting on a single unverified source with an identifiable reputational incentive and no platform corroboration.
product
OpenAI shipped a teen ChatGPT. The over-65 cohort doubled to 23% and got nothing.1 distinct publisher
build
Developer habit, priced at $965B: what Anthropic's run actually proves1 distinct publisher
leadership
BCG's new North America chair counted his own technical staff and found twice as many1 distinct publisher
science
Text watermarks land on 2 December. The detection they imply does not.1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 22, 2026