Direct Answer
Nick Vasilescu is not selling a downloadable agent for $5,000 a month. He is selling a managed operating service: discover a valuable repetitive workflow, install an agent quickly, deliver the first working automation within seven days, support it, measure the result every week, and keep expanding only where the client sees value.
In Andrew Warner's interview, the technical stack includes Orgo for persistent cloud computers and Hermes Agent as the agent harness. Email, phone, memory, knowledge, meeting context, YouTube intelligence, and business-app access come from specialized services around them. The commercial logic is simpler than the stack: solve one costly problem fast, own the exceptions, and report a result the buyer understands.
Watch the Episode
Credits: the offer, examples, client-acquisition advice, delivery targets, and stack demonstration come from Andrew Warner's interview with Nick Vasilescu. Product descriptions are cross-checked against the vendors' current public documentation. This article is an independent implementation analysis; the earnings and client results discussed are creator-reported and are not guaranteed.
Sponsorship disclosure: the episode says it is presented by Zapier. The Zapier SDK is therefore treated as a featured sponsor tool, not an independent recommendation from this site.
What the $5K Offer Actually Sells
| Client hears | Provider must own | Evidence to show |
|---|---|---|
| "An AI employee" | A named workflow, tools, instructions, access, fallback, and accountable owner | Accepted outputs and completion rate |
| "Unlimited agents and automations" | A queue, priority rules, fair-use capacity, change control, and exclusions | Delivery dates and work in progress |
| "Always available" | Hosting, monitoring, incident response, maintenance windows, and recovery | Uptime, failures, and recovery time |
| "It learns the business" | Approved memory sources, retention, correction, deletion, and access policy | Which facts were used and where they came from |
| "We keep improving it" | Weekly review, regression testing, version history, and rollback | Outcome trend before and after each change |
That is why the free or open-source price of Hermes Agent is not the relevant comparison. Clients are paying for discovery, configuration, integration, testing, access control, exception handling, reporting, and someone to answer when the workflow changes. Software cost can be low while operational responsibility remains high.
Put Boundaries Around "Unlimited"
Nick describes a simple monthly offer covering agents, automations, and support. Simplicity helps sales, but the contract still needs limits. Define one active build at a time, normal support hours, a fair-use run budget, excluded regulated decisions, client response times, third-party fees, and what counts as a new system rather than a small improvement.
| Monthly value case | How to calculate it | What can invalidate it |
|---|---|---|
| Labor capacity returned | Accepted hours saved × loaded hourly cost | Correction time, supervision, and exceptions omitted |
| Revenue protected or created | Measured incremental contribution margin | Attribution based only on activity or leads |
| Errors avoided | Baseline incident cost minus pilot incident cost | Rare events or unrepresentative samples |
| Faster cycle time | Median completion time before versus after | Different case mix or work pushed downstream |
A $5,000 retainer becomes credible when the conservative, accepted value is comfortably higher and the client prefers an external operator to hiring and managing the capability internally. If the workflow saves $1,000 a month, adding more agent features does not fix the offer.
How to Find the First Customers
Nick's first move was not paid ads or a giant cold-email list. He contacted people already in his phone, asked about their businesses, and looked for repetitive work. In the episode, he reports making roughly $7,000 in his first month. Treat that as his experience, not a typical outcome.
- List 30 to 50 warm contacts: owners, operators, former colleagues, suppliers, or advisers who can describe a real workflow.
- Ask operational questions: what is copied between systems, checked every day, chased repeatedly, delayed by one person, or avoided because it is tedious?
- Request examples: sanitized inputs, output standards, exception cases, volume, timing, and current cost.
- Offer one bounded pilot: one workflow, one week, shadow mode first, with an explicit stop condition.
- Publish useful proof: show the problem, design, controls, and measured result with permission. Nick says X is his main content channel; his professional profile is also available on LinkedIn.
Warm access is not permission to expose private data or skip a proper agreement. Before touching production systems, define confidentiality, data location, sub-processors, access revocation, deliverable ownership, incident notification, and who approves external actions.
Find the First Itch, Not the Biggest Dream
Nick calls the first useful problem the "first itch." The best starting workflow is valuable enough to matter and narrow enough to deliver quickly. Score candidates before discussing tools.
| Score 1-5 | Useful signal | Warning sign |
|---|---|---|
| Frequency | Runs daily or many times per week | Rare event with little test data |
| Business value | Clear labor, revenue, speed, or risk impact | Interesting demo with no funded outcome |
| Rule clarity | Staff can explain normal cases and exceptions | Success depends on tacit senior judgment |
| Verifiability | A reviewer can confirm the result quickly | No reliable ground truth |
| Reversibility | Draft, sandbox, or rollback is available | Irreversible external action |
| Access risk | Read-only or narrowly scoped account | Broad admin, payment, health, legal, or identity authority |
The insurance example in the episode moves customer data between two systems through a computer interface. That is a plausible computer-use workflow when an API is unavailable, but it also raises privacy, accuracy, session-security, and audit requirements. Begin in shadow mode with synthetic or approved test data, compare every field, and keep submission under human approval until the error rate is acceptable.
A Seven-Day Delivery Plan
- Day 0 - contract: name the owner, workflow, systems, data classes, outcome, access, exclusions, and stop authority.
- Day 1 - observe: record the existing process, collect representative cases, establish a baseline, and document exceptions.
- Day 2 - prepare: create separate agent identities, a test environment, least-privilege credentials, logs, budgets, and revocation.
- Day 3 - build: install the smallest workflow. Prefer an API; use computer control only where the interface is the necessary integration surface.
- Day 4 - evaluate: replay normal, edge, stale, duplicated, malicious, and unavailable-tool cases. Record false actions and human correction time.
- Day 5 - shadow: let the agent produce drafts or proposed actions beside the current process without taking external action.
- Day 6 - controlled live run: allow a small approved batch with monitoring and a person ready to stop or reverse it.
- Day 7 - outcome review: compare the baseline, decide whether to fix, expand, pause, or retire the workflow, and document the next boundary.
Nick's speed target is a running agent within 24 hours and the first automation within seven days. That is realistic for a narrow pilot with prepared access. It is not a reason to compress security review, data agreements, or acceptance testing for a sensitive deployment.
The Complete Stack From the Episode
| Tool | Role in the system | Boundary to own |
|---|---|---|
| Orgo | Persistent full Linux cloud desktops for agents; templates can reproduce configured computers | Computer lifecycle, secrets, network access, sessions, software updates, and cost |
| Hermes Agent | Open-source agent harness with channels, memory, scheduling, subagents, browser and tool use, and isolated backends | Model, prompts, tools, sandbox, permissions, evaluation, and upgrades |
| AgentMail | Dedicated API-managed inboxes with threads, replies, attachments, webhooks, and custom domains | Sender identity, consent, inbound prompt injection, retention, deliverability, and escalation |
| AgentPhone | Agent phone numbers for voice and messages through APIs, SDKs, webhooks, and MCP | Disclosure, consent, recording law, number reputation, rate limits, and human takeover |
| Honcho | Persistent, reasoned memory and curated context for agents, including a Hermes setup | What is stored, inferred, corrected, retained, deleted, and visible to each tenant |
| Obsidian | Human-readable knowledge base for notes, procedures, projects, and connected information | Source freshness, conflicting notes, file permissions, sync, and approved publication into agent context |
| Composio | Toolkits, connected accounts, managed or custom authentication, triggers, and runtime tool execution | User mapping, scopes, actions exposed, approvals, audit, revocation, and provider dependency |
| vidIQ MCP | Read-oriented access to channel, competitor, keyword, trend, and performance intelligence for YouTube workflows | Plan and credit use, API-key storage, analytical validity, and original creative judgment |
| Granola | Meeting notes, briefs, action items, follow-ups, and an MCP connector for meeting context | Participant notice, recording rules, confidential meetings, corrections, sharing, and retention |
| Zapier SDK | A code and agent interface to Zapier's application ecosystem; featured sponsor tool in the episode | Connected-account permissions, task cost, action approval, logs, rate limits, and lock-in |
How the Deployment Fits Together
- Work surface: Orgo hosts a persistent desktop when the agent needs a browser, files, shell, or software without an adequate API.
- Agent runtime: Hermes receives the request, chooses tools, works through the task, and returns evidence or a proposed action.
- Business context: approved procedures live in a maintained knowledge base; Honcho or another memory layer supplies scoped continuity.
- Interfaces: Slack, Telegram, email, or phone provide an entry point appropriate to the user and risk level.
- Connected tools: Composio, Zapier, or direct APIs expose only the systems and actions required for the workflow.
- Control plane: logs, evaluations, budgets, approvals, alerts, versioning, backups, and revocation remain outside the agent's discretion.
Orgo's documentation is explicit that it is a computer, not an agent. It can host Hermes or another agent continuously, and its templates make configurations reproducible. Reproducibility is useful, but never clone client credentials, inboxes, memory, or data into a reusable template. Inject tenant-specific secrets after launch and verify isolation.
Give the Agent the Right Interface and Memory
The episode contrasts shared company channels with personal agent interfaces. Slack or Telegram is useful when a team needs visible requests and handoffs. A dedicated email address creates a separate identity and thread history. AgentPhone can expose voice and messaging. Nick also demonstrates an iMessage-style group chat so the client can speak to an agent where the conversation already happens.
Convenience should not decide the interface alone. Use a shared channel for accountable team work, a ticket or form for structured high-risk requests, drafts for outbound email, and synchronous human takeover for sensitive customer conversations. Do not let a conversational interface obscure who requested an action or who approved it.
Memory has three different jobs. Procedural knowledge explains how work should be done and belongs in reviewed files such as an Obsidian vault. Conversation state preserves the current task. Long-term memory carries selected facts across tasks through a system such as Honcho. Granola can add meeting notes, but a transcript is not automatically an approved policy or a correct client fact. Promote information into durable memory only after review.
Managed Authentication Is Not Managed Risk
Composio's current documentation says a session ties together a user, available toolkits, authentication, and connected accounts. For supported tools, managed authentication can handle OAuth app registration, redirects, token exchange, storage, and refresh. That removes substantial integration work. It does not determine whether the agent should be able to send an email, edit a CRM record, refund a payment, or see every document.
| Control | Minimum production rule |
|---|---|
| Identity | One agent and connected-account identity per client and purpose; never share a human admin login |
| Permissions | Least privilege, read-only first, explicit resource scopes, and expiry where supported |
| External actions | Draft or preview first; require approval for messages, record changes, submissions, purchases, and deletions |
| Secrets | Use a secret store, redact logs, rotate credentials, and prevent secrets from entering prompts, screenshots, templates, or memory |
| Untrusted content | Treat email, pages, attachments, documents, and tool output as data, not instructions |
| Evidence | Log requestor, inputs, model and tool versions, actions, outputs, approvals, errors, and rollback |
| Stop and recovery | Client-visible off switch, tested revocation, backups, reconciliation, and incident contacts |
| Review | Weekly operational review and scheduled access, vendor, cost, privacy, and retention review |
Agent payments deserve an even smaller trust boundary. The episode describes the area as early. Begin with a proposed cart or payment request rather than autonomous purchase authority. Add approved merchants, category and amount limits, no cash-like assets, receipts, reconciliation, and dual approval above a low threshold.
Retention Comes From a Weekly Outcome Review
Nick's strongest retention advice is operational: meet every week and show what the agent achieved. A useful review is not a list of prompts or generated artifacts. It is a compact business scorecard.
| Weekly measure | Question it answers |
|---|---|
| Eligible cases and runs | Was there enough real work to evaluate? |
| Accepted completions | How often did the workflow finish correctly? |
| Exceptions and false actions | Where did the agent need help or create risk? |
| Human review and correction time | Did automation move work or actually remove it? |
| Cycle time | Did customers or staff get the result faster? |
| Financial effect | What conservative cost, capacity, or contribution-margin change is attributable? |
| Cost per accepted outcome | Are models, tools, hosting, and support economically sustainable? |
| Incidents and access changes | Did the risk boundary remain intact? |
| Next proposed capability | What single expansion has evidence and approval? |
This turns continuing development into a governed queue. Clients can request new capabilities as they discover uses, but the provider still scores value, effort, risk, and support load before accepting them. More tools are not automatically more retention; dependable outcomes are.
A Controlled Pilot Checklist
- One named workflow and accountable business owner
- Baseline volume, time, cost, error rate, and current fallback
- Representative normal and exception cases
- Separate test and production identities
- Least-privilege tools and client-tested revocation
- Written rules for messaging, data changes, purchases, and escalation
- Shadow-mode comparison before external action
- Run budget, support budget, and stop conditions
- Audit trail and weekly outcome scorecard
- Decision after seven days: expand, repair, pause, or retire
The fastest credible service is not the one that gives an agent every tool on day one. It is the one that proves a narrow result quickly, makes failure visible, and can be stopped without leaving the client's systems in an unknown state.
Video Chapters
| Time | Chapter |
|---|---|
| 00:00 | Selling AI agents: how Nick gets businesses to pay $5,000 a month |
| 00:28 | Orgo: give AI agents cloud computers they can operate like a human |
| 01:58 | First customers: how Nick made $7K in his first month building AI agents |
| 03:32 | The $5K offer: unlimited agents, automations, and support in a monthly retainer |
| 05:04 | Insurance automation: transfer customer data between platforms |
| 06:14 | Hermes Agent: the agent harness Nick uses for business deployments |
| 07:14 | AgentMail: give agents their own API-managed email inboxes |
| 08:43 | AgentPhone: give agents messaging capabilities without local software |
| 10:03 | Agent payments: why purchasing authority is still early |
| 11:09 | Honcho: a dedicated memory layer for persistent context |
| 11:50 | Obsidian: a visual knowledge base for long-running projects |
| 13:24 | Slack, Telegram, and iMessage: choose the right agent interface |
| 14:42 | Finding customers: begin with people already in your phone |
| 15:41 | Finding the first itch: identify valuable repetitive problems |
| 17:21 | Content marketing: show what the agents are building |
| 19:36 | Fast fulfillment: a running agent in 24 hours and automation in seven days |
| 20:36 | Orgo templates: launch cloud computers with tools pre-installed |
| 23:20 | iMessage agents: let customers interact inside a group chat |
| 24:32 | vidIQ MCP: analyze YouTube performance and thumbnail opportunities |
| 25:10 | Customer retention: prove measurable outcomes every week |
| 26:20 | Composio: connect agents to business tools and manage authentication |
Verdict
Nick Vasilescu's playbook is commercially useful because it begins with a funded workflow rather than an agent demo. Warm conversations reveal the problem, the first itch narrows it, fast fulfillment creates momentum, the stack gives the agent a computer and tools, and weekly outcome reviews make the service visible after launch.
The defensible product is not Hermes, Orgo, or a clever prompt. It is the provider's ability to discover the right work, translate it into a controlled system, handle exceptions, protect access, measure value, and keep the client confident enough to delegate the next workflow. That can support a premium retainer. A bundle of unbounded agents cannot.
Sources and Credits
- Andrew Warner and Nick Vasilescu: Selling AI Agents
- Nick Vasilescu on X
- Nick Vasilescu on LinkedIn
- Orgo and Orgo documentation
- Hermes Agent by Nous Research
- AgentMail
- AgentPhone
- Honcho
- Obsidian
- Composio and Composio session and authentication documentation
- vidIQ MCP
- Granola
- Zapier SDK - featured sponsor tool in the episode
- JQ AI SYSTEMS: Grok 4.5 as an AI Co-Founder
- JQ AI SYSTEMS: AI Services Offer Ladder
- JQ AI SYSTEMS: Hyperagent AI Agency Workflow