Agentic marketing turns market insight and execution methods into instructions an AI agent can apply itself. Growth Lab begins with the repository, prototype, URL, or idea; builds the missing product and customer context; selects relevant tools; completes the work; checks the result; and preserves evidence for the next run.
This definition is close to the practical distinction described in Ahrefs’ guide to agentic marketing: the agent selects steps and tools and continues until the job is completed. Growth Lab adds a persistent review loop around that execution.
A real run: Growth Lab markets itself
| Stage | What the agent observed or did | Inspectable artifact |
|---|---|---|
| Product model | Read the runtime, Skills, executors, and existing claims. | Repository files and live routes |
| Demand | Tested searches and decomposed 3–5 leading pages per selected query. | Dated research Memory |
| Decision | Selected tool, operating-model, day-zero, and SEO-execution intents. | Four distinct page briefs |
| Action | Created bilingual pages, technical SEO, and submitted live URLs. | Source, deployment, IndexNow response |
| Review | Established the launch baseline; performance evaluation follows. | Pending +7-day Memory |
Completed execution is recorded as completed. Market impact remains pending until evidence arrives.
The five parts of a working marketing agent
| Part | Question | Example |
|---|---|---|
| Observation | What changed in the market or product? | New search queries reached a page. |
| Decision | Which single action is supported by evidence? | Rewrite the title around the arriving intent. |
| Execution | Which tools complete the action? | Edit Next.js metadata, test, deploy, submit IndexNow. |
| Review | What outcome followed? | Compare impressions, position, CTR, and activation. |
| Memory | What should the next run know? | Date, evidence, action, result, and next recommendation. |
Start with one narrow loop
A narrow loop has a recognizable observation, an action surface the agent can reach, and an outcome that arrives within a useful time window. SEO pages, release distribution, onboarding analysis, and content repurposing can each form a separate loop.
Read Memory
→ observe one signal
→ choose one action
→ invoke the required tools
→ inspect the result
→ wait for the outcome
→ review and write MemoryEach loop benefits from specialized methods. A demand-research Skill should know how to expand keyword families and validate intent. A page-creation Skill should know how to create information gain. A review Skill should know how to remove generic content. Specialization keeps the loop concrete.
What the agent needs access to
The product
Repository access lets a coding agent see features, routes, onboarding, content, analytics hooks, and constraints in their real form. A live URL or product brief can support an earlier product.
The outside world
Search and browser access support market observation. Official APIs, exports, and authorized Clients provide structured data or external actions. Credentials stay in environment variables.
A place to remember outcomes
Memory holds dated operational evidence and next actions. The agent reads it before a new observation. The loop method remains in its Skill, where the team can inspect and improve it directly.
A 0→1 agentic marketing loop
- Open the product repository or describe the product idea.
- Ask the agent to form user and usage-scenario hypotheses.
- Collect public demand evidence around those scenarios.
- Choose one action that can create a measurable signal.
- Add or connect the smallest analytics path required to see the outcome.
- Execute, review, preserve the result, and choose what follows.
Control stays in the session
The Codex or Claude Code session is the control plane. It reads Skills, uses native browser and coding tools, invokes Clients for external APIs, edits the workspace, and collaborates with the user when access or judgment is required.
This model keeps the method readable. A team can improve the loop while using it, replace a Client, or inspect the evidence behind a decision.
Installation is a clone, and interaction is a conversation in Codex or Claude Code. There is no separate dashboard to learn, SaaS account to register, or campaign setup form to maintain.