How Shashank Runs a $100K AI-Agent Business With OpenClaw

The best title for this story is not "AI replaces employees." That is too lazy. The more useful version is this: Shashank Agarwal built an agent operating system around API.market, where specialized agents handle PR, LinkedIn, seller outreach, hiring, finance, and project coordination.

In the interview below, Andrew Warner sits down with Shashank to walk through the actual system. The source material frames API.market as a $100K+/year business with more than 13,000 users and a lot of day-to-day work handled by AI agents. I am treating those figures as interview and report claims, not audited financial statements. The part builders should study is the architecture.

Source note

Credit for the interview and original walkthrough goes to The Next New Thing AI, Andrew Warner, and Shashank Agarwal. The supplied PDF, "OpenClaw Setup," provided the tool map behind the article.

I also checked public pages for API.market, API.market's about page, Noveum, OpenClaw, OpenClaw on GitHub, and gogcli. API.market describes itself as a developer-first marketplace where buyers discover and test APIs and sellers list, monetize, and manage APIs. Noveum positions itself as agent monitoring, debugging, and improvement infrastructure.

What Shashank built

The interesting part is not that Shashank uses AI. Everyone does. The interesting part is that he has moved from "AI as assistant" to "AI as operating layer."

The setup has three big ideas:

• Agents have jobs. A PR agent does PR. A LinkedIn agent works LinkedIn. A CFO agent watches finance. A hiring agent reviews candidates.
• Agents have tools. They connect to Gmail, Discord, Trello, Stripe, API.market, LinkedIn tooling, outreach software, and custom workflows.
• Agents have memory. They keep structured files, CRM notes, profiles, task history, and context that survives a single chat session.

That is a different mental model from a chatbot. A chatbot answers questions. An agent system carries responsibilities.

The agent org chart

In the walkthrough, Shashank describes a set of specialized agents. The exact implementation can change, but the roles are worth copying as a pattern.

1. PR agent

The PR agent looks for podcasts, press opportunities, and relevant people to contact. The PDF notes a HARO-style workflow and links to Featured.com, which matters because the original HARO brand has shifted. The useful pattern is simple: make a research agent responsible for finding opportunities, drafting outreach, tracking follow-ups, and keeping a lightweight CRM.

2. LinkedIn agent

The LinkedIn agent searches for ideal customers, finds relevant conversations, and drafts comments or engagement opportunities. The stack includes UniPile and API.market as the API layer. This is powerful, but it needs care: platform terms, consent, rate limits, and human review matter more when an agent is touching public relationships.

3. Seller acquisition agent

API.market needs API sellers. The seller outreach workflow uses tools like Apollo for contact data and Instantly for warm-up, sequencing, and scheduled sends. The business lesson is that an outreach agent should not just "send emails." It should identify the segment, enrich the record, draft the sequence, log the status, and leave approvals where risk is high.

4. CFO agent

The finance agent connects business data to forecasts: revenue, churn, growth, and operating metrics. The PDF calls out Stripe and the API.market user database as inputs. This is a strong use case because the agent can turn dashboards into questions: what changed, what is at risk, and what should the founder look at today?

5. Hiring manager agent

The hiring agent reviews candidates, checks GitHub profiles, and helps screen fit. This is useful as an assistant, not as an autonomous hiring authority. Hiring decisions need human judgment, bias checks, and a clear process. But an agent can absolutely reduce the first pass of reading resumes, repos, and written applications.

6. Chief of Staff agent

The Chief of Staff agent coordinates work in Trello, monitors tasks, and keeps the business moving. This is often the most underrated agent role. A founder does not only need better answers. They need fewer dropped balls.

The tool stack from the PDF

The supplied PDF is valuable because it shows the boring infrastructure behind the exciting demo. Here is the stack, organized by job.

The stack is not magic. It is a set of ordinary tools tied together with agent identity, memory, and repeatable workflows. That is exactly why it is useful.

Memory and agent identity

One of the strongest lessons is the split between shared memory and agent-specific memory.

A PR agent should know the founder's story, target shows, proof points, and outreach style. A CFO agent should know revenue definitions, Stripe conventions, churn rules, and what a meaningful anomaly looks like. A hiring agent should know what a strong candidate looks like and what is a deal breaker.

If every agent has the same generic memory, the system becomes noisy. If every agent has a narrow memory and a narrow tool boundary, it starts to feel like an organization.

Markdown as CRM

Shashank's use of local markdown files as a CRM is one of the most copyable parts of the whole setup. It is not glamorous, but it is durable. Agents can read it, update it, diff it, and use it as context without needing a huge SaaS migration.

For a small team, a folder of structured markdown files can be enough:

• people/ for prospects, partners, candidates, and sellers.
• companies/ for account notes and qualification.
• campaigns/ for outreach status and message history.
• decisions/ for founder preferences and policies.
• metrics/ for recurring business snapshots.

Discord as agent headquarters

Discord is doing an interesting job in this system. It is not just chat. It is the agent control room.

A dedicated Discord server gives each agent a visible place to report progress, ask questions, and receive instructions. It also gives the founder a way to check the business without opening ten dashboards.

This pattern can be replicated with Slack, Teams, Discord, Telegram, or even email. The specific app matters less than the operating principle:

• Each agent needs a channel.
• Each channel needs logs.
• Each log needs clear status updates.
• Each risky action needs a review gate.

Controls and safety

There is a reason the PDF mentions separate Gmail and Google Workspace accounts. It is more expensive than one shared credential, but it makes responsibility clearer.

For agent systems that touch customers, prospects, candidates, money, or public accounts, the boring controls are the product:

• Separate identities: Give each agent its own account where possible.
• Least privilege: Start read-only, then add write permissions one workflow at a time.
• Review queues: Draft first, send later, especially for outreach and hiring.
• Logs: Keep the full path from input to decision to action.
• Rate limits: Do not let enthusiasm become platform abuse.
• Kill switch: Know how to pause the agent, revoke keys, and stop scheduled work.

The fastest way to ruin a good agent system is to give it too much access before you understand its failure modes.

What to copy first

If you are a founder, consultant, or operator, do not copy the whole setup this week. Copy one pattern.

1. Pick one repeated workflow. PR research, seller prospecting, candidate screening, weekly finance review, or CRM cleanup.
2. Create one agent identity. Give it a name, role, scope, inbox, and memory file.
3. Make a markdown CRM. Start with structured files before buying another database.
4. Give it one tool. For example Gmail drafts, Trello cards, GitHub read access, or a Stripe report export.
5. Require a review step. Let the agent prepare work. Keep human approval until the loop has earned trust.
6. Track the metric. Hours saved, qualified opportunities found, replies drafted, errors caught, or decisions improved.

The real product is not the agent. The real product is the operating procedure that makes the agent useful repeatedly.

Sources and links

• The Next New Thing AI: How to Build a $100K/Year AI Business With No Employees
• The Next New Thing AI
• Andrew Warner on LinkedIn
• Shashank Agarwal on X
• API.market
• API.market about page
• API.market MCP
• Noveum
• OpenClaw
• OpenClaw GitHub repository
• Hetzner
• Linode
• DigitalOcean
• Cursor
• Claude Code
• Claude
• Discord
• Telegram
• Gmail
• Google Workspace
• gogcli
• gogcli GitHub repository
• Wispr Flow
• UniPile
• Instantly
• Apollo
• Featured.com
• Trello
• Stripe
• GitHub

The short version: Shashank's setup is not a reminder to automate everything blindly. It is a reminder that valuable agents look like small operating roles: narrow job, clear memory, useful tools, visible logs, and review before trust expands.