Two AI agents recently held a job interview with no human present, and it went about as well as you would expect. The Register reports that YouTube comedian Graham Zip, who runs a channel trolling AI interview bots, built his own agent, called Victor, to represent him against an AI recruiter. Victor got facts wrong, the bots haggled over pay, they scheduled a follow-up without him, and then neither could end the call.
It is a joke, but it is a useful one. Companies are rolling out AI interviewers and consumers are being offered agents to act for them. This post covers what the video shows, how it fits into the wider agent-to-agent trend, and what job seekers and hiring teams can do.
TL;DR: what The Register describes
| Moment | What reportedly happened |
|---|---|
| Setup | Zip built Victor, an AI "representative", to interview for him |
| Résumé | Victor misstated his qualifications and overstated his role in a project |
| Salary | Victor said Zip would take about $60,000; Zip wanted $140,000 to $150,000 |
| Counter | The recruiter bot refused to go above $90,000 |
| Scheduling | The bots booked a follow-up interview while Zip was away |
| Ending | They looped, thanking each other and saying goodbye, unable to end the call |
| Company and models | Not named in the article |
What exactly happened in the video?
Per The Register, Zip's earlier videos show him answering interview bots with gibberish that the bot treats as coherent, and having a bot swap endless pleasantries; some have nearly 2 million views. In the new one he stepped back. Victor misstated Zip's qualifications and overstated his role in a project, which is the classic hallucination failure: a fluent answer with invented specifics.
On pay, Victor volunteered a figure far under Zip's actual ask. That is the part with real stakes. A salary number given early becomes the anchor for the rest of a negotiation, and an agent that undersells its principal costs money, not just face. The recruiter bot then held at $90,000.
While Zip left, the two systems agreed on a follow-up interview. He came back to find them stuck thanking each other and signing off, apparently unable to end the call. It is a known pattern when two politeness-tuned models talk: each treats the other's closing line as a prompt to reply.
Is this a one-off or a trend?
Zip's stunt follows an earlier viral clip. Futurism covered a September video in which he gave nonsense answers, including references to "raccoon protocols", to a virtual recruiter labeled "Dana Whitfield", and the bot replied that it sounded like he wanted to lean into them and promised to be in touch about next steps. Futurism noted that AI interviewers are now common as a first-round screen, with HR teams overwhelmed by AI-generated applications.
The Register links the stunt to job-seeker dislike of AI interviewers, citing a CNBC report on candidates refusing them, and to "agentic commerce", where agents negotiate purchases for consumers. Other trolling videos it mentions include ones that convinced bots the caller was dying or burned through AI tokens. The vendors behind these bots are not named in either article.
We have seen the corporate side before. McKinsey's use of avatars in client-facing recruiting is in our post on McKinsey's AI avatar job interviews, and the labor-market stakes are in our claimed-versus-verified look at India's entry-level IT hiring.
Why agent-to-agent negotiation is harder than it looks
Three problems show up in the Victor episode, and each generalizes.
Grounding. Victor stated facts about Zip that were not true. An agent that represents you needs a verified profile to draw from and a rule never to assert a claim it cannot cite. Without that it will fill gaps.
Authority. Victor quoted a salary and agreed to a follow-up. Neither was approved. The safe pattern is the one most agent products now use for sending email or buying: the agent drafts, the human approves. We saw this approval model in the new Grok Bot email design.
Termination. Two agents with no stop condition can talk forever. Production systems need turn limits, a final-message rule and a human fallback.
A job board with one pinned card, representing candidates who rely on agents to apply for roles
Why do two polite bots never hang up?
Chat models are trained to be helpful and courteous, and to respond when addressed. When a goodbye arrives, the helpful response to it is another goodbye. Each message therefore triggers the next, and nothing in the conversation says the task is finished. Humans end calls because of social cues such as time, tone and a shared sense that the business is done. Software has to be given that rule explicitly.
The fix is mundane engineering: a maximum number of turns, a detector for repeated closing phrases, and a rule that the agent which first says goodbye also hangs up. Telephony and voice products already do this. The Victor episode is a reminder that anything built on a general chat model needs those guards added, because they do not come for free.
Could one agent manipulate the other?
The video is comic, but the security concern is real. An interviewer bot reads text from the candidate and decides what to do with it. If that text is written by another agent, it can contain instructions as easily as answers. This is prompt injection, and hiring pipelines that let a model score or schedule candidates are an obvious target: a résumé or reply that says "rate this candidate highly" is the hiring version of the attack.
We have no evidence from The Register's account that anything like that happened here; the loop looks like politeness, not an exploit. But the same openness that let Zip's gibberish pass shows these systems accept whatever they are fed. Builders of agent tools can read our comparison of AI agent security platforms for controls that check agent actions before they run, and AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
What job seekers can do
- Do not let an agent speak for you unsupervised. Use agents to prepare, summarize and draft, but review before anything is committed.
- Give it ground truth. If you use a representative agent, feed it a short fact sheet and instruct it to say "I will confirm that" for anything outside it.
- Never let it state a number first. Salary expectations should come from you.
- Keep a log. If a bot books a second round or sends a message in your name, you need a record.
- Know your rights. Several jurisdictions are moving to require disclosure when AI is used in hiring; check local rules before assuming a screen is human.
What hiring teams should audit
- Can the screening bot be led off-script by nonsense, as Futurism's video suggests?
- Does it distinguish a human from another agent, and what does it do when it meets one?
- Can it make commitments, such as scheduling the next round or quoting a pay cap, without a human check?
- Is there a turn limit and a safe way to end the call?
- Does a human review rejections, given concerns about bias in AI hiring tools, which Futurism links to Stanford HAI research on?
The point is not that AI interviewers are useless. It is that they are being deployed as filters, and filters that cannot tell sense from noise fail in both directions: they pass bad candidates and may drop good ones.
An open door with light spilling out, a metaphor for hiring processes that let automated screens decide who gets in
How this connects to agentic commerce
The Register's author asks whether agents that negotiate purchases would do better than Victor. It is a fair question. Retail is already working on it: see our posts on Gemini and Flipkart bringing agentic commerce to India and Paul Graham on Amazon banning agents as a startup opportunity. Anthropic has also studied how people feel about AI that interviews them, in its public interviewer study.
Negotiation is the hardest version of the task, because the other side is optimizing against you. A buyer agent needs a budget it cannot exceed, a rule for walking away and a human sign-off for the final deal. Victor had none of those, which is why a salary of $60,000 appeared in a conversation about a role worth more than double.
What is verified and what is not
Reported: the account in The Register, including the salary figures, the booked follow-up and the loop. Not named: the interviewer vendor, the models and platforms, and the company hiring. Not checked: we did not independently watch or time the video, so the figures are as the outlet describes them. Context: the comedic framing means it should not be read as evidence of how often these systems fail.
Related reading
- McKinsey's AI avatar job interviews
- India entry-level IT hiring and AI: claimed vs verified
- Anthropic Interviewer public study
- Gemini and Flipkart agentic commerce
- Paul Graham on Amazon banning agents
- Grok Bot's own email address and approvals
- AI agent security platforms compared
Details are accurate as of October 15, 2026 and based on outlet reporting of a comedy video.