In this walkthrough, Andrew Warner opens the agent chats behind the work: promotion ideas, rent and utilities, Marketplace listings, inbox triage, physical mail, contacts, and an anniversary. The useful question is not whether one bot can do all of that. It is which steps benefited from judgment, which should have been simple automation, and where Andrew still had to check the result.
Andrew's strongest examples pair a narrow agent job with an existing system of record and a clear handoff. Peri turns actual conversations into promotion ideas; Maggie prepares tenant invoices; Postman presents mail decisions; Dex backs up contacts before merging. The most instructive failure is an agent repeatedly polling for events and running up extra usage. Andrew moved that work to Zapier. The agent-made slide deck also missed a privacy redaction, reminding us that a polished presentation is not a completed review.
Watch the Video and Explore the Deck
Credit and disclosure: These are Andrew Warner's workflows and outcomes, not tests run by this site. The video is presented by Zapier. Andrew also shares a 51-slide episode deck with agent-chat screenshots. It is useful evidence of his process, though the deck itself is an agent-produced artifact and Andrew points out a missed redaction.
Ten Jobs, Grouped by Outcome
The headline says ten agents, but the video mixes named specialists, supporting automations, and multiple jobs done by one specialist. These are ten jobs Andrew demonstrates or describes, not a claim that he deployed ten isolated bots.
1. Find a genuine publicity angle
What happened
Peri looked through Andrew's authorized email, Slack, and video conversations for accomplishments or topics he actually cared about. It proposed three ideas a day. Andrew rewrote and published one how-to post; he reports a high-performing follow-up after a prominent X repost.
Practical limit
The agent found the angle, but Andrew supplied the opinion, edited the draft, and chose what to publish. Do not give a PR agent broad access to private conversations or let it post without consent and claim checks.
2. Save testimonials with their context
What happened
Peri noticed a guest's comment about using an SEO tool after an interview and filed it in Notion as potential proof for future guest or sponsor conversations.
Practical limit
A positive comment is not permission to turn a person's words into an unrestricted endorsement. Keep the original link, wording, date, and usage permission with the record.
3. Spot an unlabelled rent payment
What happened
Maggie noticed a wire matching the rent amount even though its memo did not identify it cleanly. Andrew says he otherwise needed to call his bank to establish what it was.
Practical limit
An amount match is a lead, not proof of payer or invoice settlement. Verify the transaction against the bank record and tenant before marking rent paid. Financial access should be minimal and read-only where possible.
4. Prepare auditable utility invoices
What happened
Maggie converted a garbage bill into a tenant invoice in Notion and calculated a partial-month electric charge from the utility's daily usage and rate information, making the calculation visible.
Practical limit
Check lease terms, dates, rates, taxes, and the source bill before issuing anything. A transparent calculation can still use the wrong obligation or rate. Keep a person responsible for final invoices.
5. Replace agent polling with triggers
What happened
Andrew shows several extra-usage charges after agents repeatedly checked for events. He moved predictable waiting and routing to Zapier, with an agent called Zapier looking for more such migrations.
Practical limit
This is Andrew's cost experience, not a universal savings benchmark. Zapier's Next Gen Zaps are marked early access on its site. Compare total cost per completed task, including errors, AI calls, and human review.
6. Relist and negotiate on Marketplace
What happened
A Marketplace agent fixed a trailer listing, followed up with interested buyers, negotiated within Andrew's price floor, tracked offers, and helped sell other items, including a mower.
Practical limit
Andrew authorized more autonomy than many sellers would want. Before enabling replies, define the permitted items, truthful claims, response frequency, price floor, meeting safety, and when a person must approve a commitment.
7. Surface urgent mail; archive the rest
What happened
Work Mail Butler presents Andrew's important messages each morning. He expanded its permission from suggesting archives and unsubscribes to performing those lower-stakes actions and reporting them afterward. It flagged a refund request for him.
Practical limit
Archive and unsubscribe are not equally reversible in every service. Use allowlists, a reviewable action log, and special handling for clients, bills, legal notices, and refund requests. Never publish inbox screenshots without manual redaction.
8. Route physical mail
What happened
PhysicalAddress.com receives mail and shows scanned envelopes. Postman asks Andrew whether to open, forward, shred, or ignore each item. For a vehicle-registration letter, Andrew approved an email inquiry; the reply clarified that the fee had already been paid.
Practical limit
A mailed notice can carry deadlines or sensitive data. Keep the original, verify dates and sender, and require approval before shredding, forwarding, or contacting authorities. The agent did not independently resolve a legal obligation.
9. Enrich and deduplicate contacts
What happened
Wally used message context to identify unknown WhatsApp contacts. Dex backed up Andrew's address book before proposing merges and filling missing names from available context. Andrew describes thousands of records and a large duplicate cleanup.
Practical limit
Guessing identity from private conversations can be wrong and exposes third-party data. Export first, test five records, review ambiguous matches, and keep a restore path. The deck also showed a name Andrew thought should have been blurred.
10. Plan a personal experience
What happened
An anniversary agent collected gift and experience ideas, learned from Andrew's corrections, sent reminders, and led him to a booking page. Andrew completed the purchase himself to choose the right payment card.
Practical limit
Useful research and nudges do not justify unattended purchases or sharing a partner's private preferences widely. Keep payment, booking, and personal messages under human control.
The Cost Lesson: Let Code Wait, Let Agents Judge
Andrew shows extra charges appearing after he opted into pay-as-you-go use. His explanation is specific: agents waiting for calendar or other events consumed usage without producing new decisions. An event-triggered workflow can wait cheaply and call an agent only when a new item needs classification, writing, or exception handling. Zapier describes this split as deterministic steps plus AI where needed, while marking Next Gen Zaps as early access. It is not proof that every Zapier migration saves money.
A workable comparison is cost per accepted task: platform fees + model usage + review time + error repair. Test it on one workflow for a week. If a fixed rule can route the event correctly, keep the model out of that step.
The Mistakes Are Part of the Demo
Andrew asked the deck-producing agent to blur sensitive information, yet he spotted a refund detail and later a contact name that had not been handled consistently. He says he will ask a human editor to repair the slide. This is a more useful lesson than declaring that agents never make mistakes: human review belongs at the point of publication or irreversible action, with feedback logged so the next run improves.
The official Grok Bot product page says bots can use a persistent cloud computer, sign into tools, run routines, and request approval. It also says bots for one user share that computer and its files and logins. Treat that as a shared access boundary, not ten sealed workspaces. Start each role with the smallest permissions it needs; audit external sends, purchases, invoices, financial changes, contact merges, and destructive mail actions separately.
What I Would Copy First
Start with the least sensitive version of Andrew's strongest pattern: a weekly proof and publicity brief using only public posts and material you explicitly approve. Let a simple trigger collect links, have an assistant propose three useful angles, and make a person choose, edit, and publish one. Measure time saved and whether the resulting post generates relevant conversations, not just impressions.
Only after that works should you consider inbox, property, contact, or payment access. The deck is excellent for seeing the workflow, but it is not a permission template for someone else's accounts.
Jump to the Main Demonstrations
| Time | Topic | Time | Topic |
|---|---|---|---|
| 00:17 | Peri: publicity | 04:07 | Maggie: rent and bills |
| 06:30 | Extra usage and Zapier | 08:35 | Marketplace manager |
| 11:20 | Work Mail Butler | 12:13 | Mistakes and feedback |
| 13:42 | Postman: physical mail | 15:23 | Wally: WhatsApp contacts |
| 17:21 | Dex: backup and merge | 18:48 | Anniversary planning |
| 20:53 | Research deck and editing | 21:32 | Feedback loop |
Business Idea: A Reviewed Agent Workflow for One Niche
The commercial opportunity is not “ten bots for everyone.” It is one well-defined job with a trigger, a trusted record, an exception path, and a human-approved result. This prompt turns Andrew's examples into a bounded service test that works with whichever tools you already have.
Design one agent service
Choose the task, control the access, and measure the real cost.
Sources and Evidence Limits
Primary account: Andrew Warner's video and his episode deck. Product references: Grok Bot, Zapier, Facebook Marketplace, and PhysicalAddress.com. Andrew's screenshots and narration show what he says happened in his accounts; we did not independently audit the transactions, agent logs, total costs, or privacy controls. Features, access, and prices can change. The video is sponsored by Zapier.