Marketing Agents Masterclass: Two AI Systems to Grow Your Startup

Marketing agents are the new coding agents. Coding agents made it dramatically easier to build software. Marketing agents give us a way to build the systems that bring customers to that software.

But “build me a marketing agent” is about as useful as “do marketing for me.” You need a specific job, a clear process, the right data, and the infrastructure to keep it running.

There are two systems worth understanding here. The first finds potential customers through LinkedIn engagement, enriches their contact information, and runs outbound across email and LinkedIn. The second turns real conversations inside your company into an organic LinkedIn content engine for your entire team.

The important part is not some magical prompt. It is knowing who you want to reach, why they might care, and which parts of the work should be ordinary software versus an LLM making a judgment.

Table of Contents

• Start With the Right Definition of a Marketing Agent
• Agent One: Signal-Based Cold Outbound
• Compliance Is a Separate Part of the System
• Deploy the Workflow, Not Your Coding Session
• Agent Two: An Organic LinkedIn Content Engine
• Earned Media Is a Distribution Asset
• The Real Skill Is Designing the Process
• Get Help Deploying These Marketing Agents

Start With the Right Definition of a Marketing Agent

People get hung up on the difference between an agent and an automation. For this kind of marketing work, the useful definition is simple:

A marketing agent is code, connected to a live data stream, with an LLM attached where judgment is needed.

A daily scheduled job does not need to think. Neither does an API call that fetches a list of posts. But deciding whether a company fits your ideal customer profile might require reasoning. So might answering a prospect’s question or turning a customer objection into a useful social post.

This distinction matters because you should not be paying an LLM to rediscover how to make the same API call every day. Pay the coding agent to build the software that makes that call. Then run the software on inexpensive compute.

Think about a good media buyer. They research creative angles, make new ads, test them, remove losers, and promote winners. The opportunity is to encode that process, not hand an unrestricted model an ad account and hope it figures everything out.

The same principle applies to outbound and organic content. Build a repeatable process first. Add intelligence at the decision points.

Agent One: Signal-Based Cold Outbound

Cold outreach gets harder when every inbox is full of generic AI-generated messages. More volume is not a particularly interesting answer to that problem.

A better answer is better timing and better targeting.

Traditional outbound starts with relatively static information: job title, company size, industry, and location. Those details still matter. But LinkedIn engagement adds another layer: someone is interacting with a topic relevant to your product right now.

A like or comment is a signal of interest. It is not proof that someone wants to buy, and it is not permission to contact them. But it gives you a much stronger starting hypothesis than a job title alone.

The outbound system follows this sequence:

1. Monitor relevant creators and company pages on LinkedIn.
2. Find new posts and extract the people who engage with them.
3. Check whether those people and their companies fit your target customer profile.
4. Find contact details through waterfall enrichment.
5. Verify email addresses before sending.
6. Run email and LinkedIn outreach through dedicated platforms.
7. Manage replies and track whether prospects book a call.

1. Find the Creators Your Customers Already Follow

Start manually. You do not need an agent to automate every five-minute research task.

Build a spreadsheet of roughly 10 to 20 creators or company accounts whose content attracts your ideal customers. Many businesses already know who these accounts are. If not, use LinkedIn search and your own feed to find them.

The question is not simply, “Does this person talk about my category?” It is, “Would my target customer interact with this content?”

For example, a search around AI for WordPress might sound perfectly aligned with an AI WordPress product. But if the results do not attract the right people or contain meaningful engagement, the search is not useful. A broader search around AI marketing might surface a much stronger audience for a marketing product.

A post demonstrating an AI video-editing workflow could be valuable because the people engaging with it may be marketers evaluating similar tools. A company page such as Clay can work just as well as an individual creator. You care about the audience, not whether the account belongs to a person.

Cody’s working heuristic is that a handful of outlier accounts can expose you to a large share of a niche’s relevant audience, roughly 80% in the way he frames it. Treat that as a sourcing heuristic, not a guaranteed measurement.

The point is diminishing returns. Once you have enough relevant lead volume, adding another hundred marginal accounts may create more noise than value.

This also solves a related problem in creative work. Agents left in a closed loop tend to repeat themselves. Tracking real human creators gives the system fresh hooks, formats, and topics to learn from.

2. Use Apify to Extract Posts and Engagers

Apify Actors are tools for collecting and processing data. In this setup, Apify supplies the scraping layer, while Claude Code or Codex helps build the software that calls it.

The LinkedIn Actors used in Cody’s workflow come from API Maestro. Three functions matter:

• Profile posts: Retrieve posts published by a source account.
• Post reactions: Retrieve people who reacted to a post.
• Post comments: Retrieve people who commented on a post.

Actor maintenance matters. A marketplace can have plenty of options, but an integration is only useful if it keeps working. Choose maintained tools rather than assuming every scraper is equally reliable.

The local development process is straightforward: provide the coding agent with the API documentation, access to the API key, and a LinkedIn post URL. Have it build and run a script that extracts the engagers and deduplicates their public profile URLs.

In the example, one post produced 63 raw engager records before the workflow finished processing them. That illustrates the input volume, not a promise that every record becomes a qualified lead.

Once the local script works, move it to a server and schedule it to run daily:

1. Check the source accounts for new posts.
2. Retrieve the new post URLs.
3. Extract reactions and comments.
4. Deduplicate the resulting profiles before continuing.

Now you have a recurring source of potential customers rather than a one-time list.

3. Qualify People Before Paying to Enrich Them

Engagement is only the first filter. Someone can love your category without being a customer you can serve.

Before enrichment, have the system research the person and their company. Relevant questions might include their role, what the business does, and how many employees it has.

Then compare those findings with your ideal customer profile, or ICP.

This is a useful place for an LLM. It can interpret messy company information and decide whether the evidence matches your criteria. Profiles that fit move forward. Profiles that do not fit should not consume the same enrichment budget.

Interest tells you where to look. Customer fit tells you whether to proceed.

4. Find Contact Details With Waterfall Enrichment

Waterfall enrichment means checking multiple data providers in sequence, passing only the unresolved records to the next provider.

The stack discussed here starts with GitLeads, moves to Apollo, and then uses options such as Origami or Prospeo for remaining gaps. LeadMagic is another option Cody uses particularly for mobile phone numbers.

Here is the example with 50 LinkedIn profiles:

• First provider: GitLeads finds 32 email addresses.
• Second provider: Send only the remaining 18 profiles to Apollo, which finds another 10.
• Later providers: Send the final eight unresolved profiles to another service.

You are not paying every provider to process every person. You are progressively filling the gaps.

The ordering should reflect your own economics: start with a provider that gives you a good combination of cost and accuracy, then move toward more expensive options when necessary. You can keep extending the waterfall, but the last few matches may not justify the extra spend.

Some services, including Origami, can aggregate the waterfall for you. Instead of managing each provider separately, you submit a profile and let the service work through its available sources.

Contact coverage varies. A LinkedIn URL gives you a useful lookup key, not a guarantee that an email address or phone number exists in a provider’s database.

5. Verify Emails Before Sending

Finding an email address is not the same as confirming it is safe to send to.

Run the results through a verification service such as MillionVerifier. The output helps distinguish valid addresses from invalid, risky, or catch-all results.

Cody specifically recommends a second verification step for results from GitLeads and Apollo rather than relying entirely on the original provider’s checks.

Why bother? Invalid addresses create bounces, and bad sending behavior can damage deliverability. For this workflow, the cleanest rule is to send only to addresses verified as valid.

6. Separate Your Outbound Email Infrastructure

Do not run cold outreach through the same domain your team relies on for normal business.

The suggested setup separates four email functions:

• Cold outreach: Dedicated sending domains and inboxes.
• Email marketing: Newsletter and marketing communications.
• Transactional email: Product messages such as password resets.
• Business email: The accounts your team uses to run the company.

The reason is risk separation. If cold sending develops reputation problems, you do not want password resets or normal customer conversations suffering alongside it.

Options for purchasing and hosting outreach inboxes include Inbox Kit, Instantly’s prebuilt inbox offering, and Hypertide, a partner Cody works with. You buy the domains and pay to host the inboxes, then connect them to sending software.

For a rough budget, Cody describes about $100 per month in inbox infrastructure for approximately 10,000 cold email sends, based on his partner setup. He also cites an Instantly tier around $97 per month.

That puts the sending software and inbox starting point around $200 per month. These are illustrative figures, not a current quote or an all-in operating budget. Domains, scraping, enrichment, verification, hosting, and model usage can add costs.

7. Add LinkedIn Outreach

The same qualified audience can feed a LinkedIn outreach workflow. HeyReach and Botdog are two tools discussed for LinkedIn messaging campaigns, and both expose APIs.

LinkedIn InMail is another route Cody has seen teams use successfully with this targeting strategy.

The important architectural point is that the messaging platform handles delivery, while your software handles the upstream research, qualification, and preparation.

Technical access does not remove platform restrictions. Review LinkedIn’s rules on scraping and automated activity before putting a workflow into production.

8. Connect Replies to an Inbox Agent

Sending is only half the job. The more interesting system manages what happens after someone responds.

Instantly’s API can support account management and campaign inputs. Your software can generate personalized copy outside the platform and push the resulting text or variables into it.

Webhooks provide the return path. When a positive reply arrives, the platform can notify an agent running on your server.

Give that agent the context it needs: product information, relevant answers, and a clear objective such as helping an interested prospect schedule a demo.

The workflow can then:

• Answer questions using the supplied business context.
• Continue the conversation toward a relevant next step.
• Share a scheduling link.
• Check Calendly or Cal.com to see whether a meeting was actually booked.
• Support later follow-ups for appropriate prospects who went quiet.

A six-month re-engagement cadence is one example Cody gives. That should still respect opt-outs and the rules governing your outreach.

This is the “SDR in a box” idea: source prospects, qualify them, find contact details, prepare messages, send through dedicated platforms, and handle the conversation afterward.

It does not require an elaborate agent framework. For a narrowly defined job, a small service with API integrations and an LLM can be enough.

Compliance Is a Separate Part of the System

Being able to obtain contact information does not automatically mean you can use it however you like.

Data sourcing, privacy obligations, platform terms, email outreach, and phone outreach are separate considerations. Rules differ across jurisdictions, including between the United States and the European Union.

For US commercial email, the FTC’s CAN-SPAM compliance guide is a useful starting point. Do not treat a purchased address, a public profile, or an engagement signal as blanket authorization.

This is not legal advice. Get the requirements clear for your market before you automate sending. Compliance belongs in the workflow design, not in a cleanup project after launch.

Deploy the Workflow, Not Your Coding Session

Claude Code or Codex helps you build the system. It does not need to remain the runtime for every repetitive task.

If the workflow runs locally, you can often deploy the resulting code to a server and run it hourly, daily, or in response to events. Railway is one hosting option discussed here.

A server is simply a computer somewhere else that stays on. It runs the code while your laptop is closed.

The broader infrastructure has three jobs:

• Data pipeline: Bring fresh information into the system.
• Data storage: Keep the records and context the workflow needs.
• Runtime: Execute scheduled jobs, receive webhooks, and make decisions.

A warehouse such as ClickHouse can be part of that setup. The exact stack matters less than making sure the software has the data and access required to do its job.

This is the useful version of “marketing is code.” A creative workflow may involve research, a script, and calls to image or video-generation services. An outbound workflow involves profiles, enrichment, sending, and replies. Both become systems you can build and improve.

Use inference when you need interpretation or creation. Use ordinary code when you already know the steps.

Agent Two: An Organic LinkedIn Content Engine

Not every company wants cold outbound as its primary growth motion. The second system builds organic distribution across the team.

Imagine a seven-person sales team posting useful, original LinkedIn content every day. The hard part is not connecting seven accounts to a scheduler. It is finding enough worthwhile ideas without making everyone sound identical.

The answer is to start with the knowledge already inside the business.

1. Mine Human Experience for Ideas

A weekly conversation is enough to create the input. Ask each person what they learned, what surprised them, and what came up repeatedly in customer conversations.

This works for salespeople, technical teammates, and other people close to the work.

You can also use existing source material:

• Recorded sales calls and Gong conversation records.
• Relevant internal Slack discussions.
• Notion documents and internal notes.
• Technical discussions and codebase context.
• Podcasts and other long-form conversations.

A prospect explaining why they did not buy can be an excellent starting point. That objection might reveal a misconception, a trade-off, or a problem other buyers are dealing with.

Compare that with asking a model to “write good LinkedIn content.” Without fresh source material, you are likely to get something generic. More output does not fix the absence of an original idea.

External material can also inspire a post. The useful move is to add your own interpretation and angle, not pretend someone else’s experience happened to you.

2. Extract Insights and Turn Them Into Posts

Once you have source material, use an LLM to identify the ideas and draft posts. Cody mentions Claude Sonnet as a model that can be sufficient for this writing step.

The workflow is not “invent expertise for seven people.” It is “help seven people communicate the expertise they already have.”

That distinction is what gives the content a reason to exist. Salespeople hear objections. Engineers encounter implementation problems. Founders make trade-offs. Those are more useful inputs than a blank prompt asking for thought leadership.

3. Schedule Across Team Accounts

Ordinal is the scheduling tool used in this setup. It supports multiple connected LinkedIn accounts and provides API or MCP access for scheduling posts to individual accounts.

That lets the system move from source material to drafts to scheduled distribution without manually copying every post into every profile.

The opportunity for the person managing this is scale. Instead of being responsible for typing and publishing one account’s posts, they can manage the content process across many accounts.

4. Feed Performance Data Back Into Writing

Ordinal also provides analytics across connected accounts and individual posts. Feed that data back into the system so the next writing cycle has evidence about what worked.

Which topics earned impressions? Which angles attracted attention? Which ideas deserve another version?

Cody uses the words “snowball” and “remix” when directing the next round of creative. They communicate the right behavior: develop a successful idea further instead of starting from nothing every time.

This mirrors what a strong social media manager already does:

1. Prospect for ideas.
2. Create content around those ideas.
3. Publish.
4. Read the performance data.
5. Bring successful ideas back into the content calendar.

Cody describes revisiting proven ideas on roughly a 90-day cadence. That is his practice, not a universal formula. The underlying lesson is to build a library of ideas that resonate, then keep finding useful ways to express them.

You are prospecting for winners. Once you find one, you do not need to abandon it just because it worked.

Earned Media Is a Distribution Asset

Organic content is not just a personal-brand exercise. It creates repeated exposure to the people who might eventually buy from you.

One way to understand its value is to compare it with paid distribution. Cody uses an illustrative LinkedIn advertising cost of about $22 per thousand impressions. On that basis, 1,000 organic impressions represent roughly $22 worth of comparable paid exposure.

That is a way to frame earned media value, not cash in the bank and not proof that organic and paid impressions convert equally.

Some platforms can also pay creators directly, which creates an interesting dynamic: content can build lead pipeline while generating platform revenue. That does not apply to every account or every platform, but it changes how you think about investing in distribution.

You Do Not Have to Be the Face of Every Account

If you do not want a personal brand, you still have options.

You can help teammates develop their own voices. Or you can build a theme-based page around a subject your customers care about.

Growth Tactics, associated with Julian Shapiro and Demand Curve, illustrates the idea. People follow useful growth content, then discover the business behind it.

A topic page can aggregate, organize, and interpret information without making the founder the central personality. The value still has to be there. An anonymous page publishing generic filler is not a distribution strategy.

The model is simple: build an audience around an existing interest, then connect that interest to a relevant product or service.

The Real Skill Is Designing the Process

The social media role is evolving toward managing systems: sourcing ideas, preserving useful human context, coordinating accounts, and improving the process from performance data.

The same is happening in outbound. The leverage comes from connecting targeting, qualification, enrichment, sending, and replies into one coherent workflow.

Both systems start with market evidence. For outbound, that evidence is relevant engagement plus customer fit. For organic content, it is real conversations plus audience response.

This is also how to think about building a business. Instead of trying to force the market to want your idea, ask what people already want, whether you can build it, and whether you can reach them.

Marketing agents do not remove the need for that judgment. They let you turn it into a system that keeps doing useful work.

Get Help Deploying These Marketing Agents

For fast-growing companies that want implementation help, Cody’s company, Graphed, offers a platform for these workflows along with forward-deployed software engineers who build the implementations on it.