AI marketing agents: the 5 levels of agentic marketing, and how to climb each one
Updated
Agentic marketing is marketing work run by AI agents that plan and execute tasks, not by a person typing prompts. It climbs in 5 levels: generic prompts while doomscrolling, detailed skill docs, skills tied to one brand foundation file, agents with a live data layer that run workflows on their own, and agent fleets that share memory and learn from each week's results. The hardest jump is level 2 to 3: writing your taste down for the first time.
- A 4-question self-audit that places you on the ladder
- Each level's routine, ceiling and next jump, with a skill skeleton and brand foundation template
- A Claude Code orchestration prompt for a 4-agent research workflow
What are the 5 levels of agentic marketing?
The 5 levels of agentic marketing run from a marketer doing everything at the keyboard to agents that ship work and learn from results while the team sleeps. Most marketers have every AI tool open and still no compounding output, because they don't know which level they're on or what the next move is.
| Level | What it looks like | The ceiling |
|---|---|---|
| 1. Doomscrolling + generic prompts | Scroll for research, paste one-line prompts into a chatbot | Output tops out at your keyboard hours |
| 2. Detailed skills + swipe files | 100-300 line skill docs per task, a swipe file by brand and pattern | Your time: every piece still needs you |
| 3. Skills + brand foundation | One foundation file every skill reads | You still run skills in sequence by hand |
| 4. Agents + intelligence layer | Type a goal; agents plan, pick skills and execute on live data | Subagents work in isolation, with no memory |
| 5. Autonomous agent fleets | Agents share context and memory, learn from real engagement | None worth naming yet |
Which level of AI marketing agents are you on?
Answer these four questions in about 30 seconds. Your most frequent answer is your level.
| Question | Level 1-2 | Level 3 | Level 4 | Level 5 |
|---|---|---|---|---|
| Where does the first idea for a brief come from? | Scrolling TikTok or Instagram (1); a Notion swipe file (2) | A brand foundation file plus last week's winners | An agent surfaces a shortlist of proven ideas to pick from | Research to ideation briefs written overnight, ready to audit |
| Hours a day YOU spend at the keyboard on marketing | 4+ | 1-2 | Under 30 min | Roughly zero |
| You take a week off. What happens? | Output stops | Limps on prebuilt skills | Agents run as scheduled | The fleet ships and feeds next week's brief generator |
| A junior marketer joins tomorrow. What do they do? | Replace you at the keyboard | Run your skills against new inputs | Approve agent outputs | Audit fleet edge cases and feed memory |
How do you move past level 1 generic prompts?
Move past level 1 by turning your 3 most repeated tasks into detailed skill docs, which takes one afternoon. At level 1 you open TikTok "for research" at 8am, screenshot hooks into a doc nobody reopens, paste one into a chatbot for 10 versions and get generic results because the model has no memory of your brand. By 2pm one piece ships. Output tops out at your keyboard hours, and with no memory the work never compounds.
This is also the role most at risk: a16z's Growth Engineer Fellowship trains 65 people in an 8-week cohort to do it with agents. Pick 3 tasks (subject lines, X hooks, short-form video hooks) and go from a 1-line prompt to a 50+ line skill doc with frameworks, examples, voice rules and edge cases:
SKILL: [name the workflow]
GOAL: what this skill produces in plain English
PROCESS: numbered step-by-step for the agent
TOOL: which model/IDE this runs through
KNOWLEDGE: links to brand voice, ICP, banned phrases, last 5 winners
EDGE CASES: 2-3 weird-input scenarios + how to handleFor Claude-ready examples, see Claude skills for marketing.
What holds marketers back at level 2?
Time holds level 2 back. Detailed skills for email, landing pages and cold DMs give you clean copy, but every piece still needs you to open a skill, paste inputs, audit and paste into the next one, so the skills multiply you instead of replacing your hours. Each skill is also an island: the email skill doesn't know what the landing page skill knows.
The next jump (one weekend) is a brand foundation doc. Audit your existing skills first, because broken skills carry their problems up a level:
- Each has a clear GOAL / PROCESS / TOOL / KNOWLEDGE section
- Each has 3 examples of good output and 3 examples of bad output
- Each has a banned-phrases list
- Each names a specific ICP ("DTC supplements $1M-$5M ARR", not "marketers")
- Each is under 500 lines (longer means you're hiding context; split it)
To fill skills with the words your audience really uses, pull them from real videos and comments with social listening.
What goes in a brand foundation file for AI agents?
A brand foundation file holds your positioning, voice rules, banned phrases, your dialect and your audience's, customer quotes, ICP segments, competitors, receipts and hard bans, in one doc every skill reads. With it, 12 pieces of copy read out loud all sound like you. The ceiling: you are still at the keyboard running skills one after another.
# [Brand name] - brand foundation
## 1. Who we are
- One-line positioning (what / who / outcome)
- 3 things we sell against (the legacy way)
- 3 things we believe that the market doesn't yet
## 2. Voice rules
- Register per channel (casual / professional / hybrid)
- Sentence rhythm (fragments / mountain copy / long-form)
- ALL CAPS rules (when, where, never full sentences)
- Lowercase rules per channel
- "I" vs "we" vs "the team" rules
## 3. Banned phrases
- Em-dashes, "It's not X. It's Y." patterns
- LinkedIn-growth-carousel cliches
- Corporate-press-release voice
## 4. Words we use constantly (the dialect)
10-20 words your brand uses religiously.
## 5. Words our audience uses (their dialect)
Mined from sales calls / forums / DMs.
## 6. Customer quote bank
20-50 verbatim quotes, tagged by ICP + emotion + type.
## 7. ICP segments (top 3-7)
Per segment: real titles, top 3 pains verbatim, tribal phrases, write-when triggers.
## 8. Positioning against alternatives
Per competitor: what they do well, where they fail, verbatim complaints.
## 9. Case studies + receipts
Named real people + verifiable results.
## 10. Hard bans + edge cases
Things that never ship, plus cases that LOOK like bans but aren't.One foundation feeds every skill, so consistency becomes automatic. Section 5 is the hardest to fill by memory; the viral hooks data shows how audiences in your niche phrase things.
How do AI marketing agents run a workflow on their own?
At level 4 you type a goal in plain English and an orchestrating agent plans the steps, picks the skills and runs them on live data while you work in another tab. A typical stack has four agents, and your morning starts at "here are today's 12 winners with transcripts, scores and briefs attached" instead of opening TikTok. Keyboard time drops by about half.
| Agent | Job |
|---|---|
| Research agent | Pulls raw data, removes duplicates, filters by view floor and outlier ratio |
| Hook agent | Maps hook patterns and clusters them by archetype |
| Brief agent | Writes the batch of briefs, referencing the brand foundation |
| Critic agent | Scores everything and flags banned phrases or off-voice lines |
The shift is from human in the loop to human over the loop. Paste this into Claude Code, with your data layer connected (Virlo's TikTok MCP and Instagram MCP provide the outlier data):
GOAL: Run my [weekly viral video research workflow] end to end
and deliver output to Slack #content-pipeline.
PROCESS:
1. Research agent pulls top 20 viral [TikTok/IG] videos
in [niche] from last [7 days]. Filter: views > [floor],
outlier ratio > 2.0x.
2. Hook agent maps each video's archetype (curiosity gap /
identity callout / contrarian / proof drop / pattern
interrupt / pain agitation / story / question).
3. Brief agent (references brand foundation at [link]) drafts
12 briefs based on top patterns.
4. Critic agent scores each: hook strength 1-10, brand voice
match 1-10, estimated engagement.
5. Compile as ranked table, drop in Slack, save raw to Notion.
KNOWLEDGE: brand foundation [link], skill docs [link],
last week's winners [link], ICP one-pager [link].
EDGE CASES:
- <10 videos pass view floor → lower 25% and re-run
- Brief fails brand voice check → flag, skip, don't auto-fix
- Empty data substrate → notify and waitThe ceiling: the subagents never talk to each other, you still stitch their outputs together, and run 1 and run 2 share no lessons.
What does a level 5 autonomous agent fleet look like?
A level 5 fleet is agents that share context: research findings feed the writer, critic feedback loops into the next version, and last week's performance shapes this week's briefs. Memory is what wins. By run 20 the system holds audience sentiment, patterns per segment and a record of which hook archetypes converted, so marketing becomes infrastructure that ships while you sleep.
Matt Berman (@TheMattBerman), who runs a New York marketing agency, built a Monday Drop system on marketing agents doing daily niche research per client. Every Monday at 7am, before standup, it has already pulled the prior week's outliers, mapped hook patterns, surfaced branded concepts per client and posted them to Slack. It did 11 million views in 28 days. The setup is documented in the Monday Drop install kit.
The common stack is Claude Code orchestrating against a data layer, with scheduled tasks running overnight. Anthropic's own growth function (one non-coder, single seat) ships out of a stack like this. To start with a single always-on agent instead of a fleet, see Virlo's marketing agent.
How long does it take to climb each level?
Climbing from level 1 to level 4 takes a few weeks of focused work; level 5 is ongoing.
| Jump | Time cost | What you build |
|---|---|---|
| Level 1 to 2 | One afternoon | 3 detailed skill docs (50+ lines) |
| Level 2 to 3 | One weekend | 1 brand foundation doc |
| Level 3 to 4 | 1-2 weeks | 1 workflow automated end to end |
| Level 4 to 5 | Ongoing | Memory, cross-agent context and scheduled tasks |
The hardest jump is level 2 to 3, because it is the only one where the work is writing your own taste on paper for the first time. Once it's written, the rest is mechanical. The gap between level 1-2 and level 3-5 marketers comes down to a brand foundation an agent can read: agents are a commodity, but years of judgment baked into a file are not. Like a film director who never touches the camera, you stop writing and start deciding what stays.
Give your agents a data layer: 23.3M viral TikTok and Instagram videos, with hooks, formats and outlier scores they can read.
How this resource was made
- Source
- The original Virlo resource "The 5 Levels of Agent-Driven Marketing" by co-founder Jaife Esienna (self-audit, levels, templates, orchestration prompt and cost table).
- How this page is made
- Written by the Virlo team from the original resource; templates and the orchestration prompt kept word for word apart from the platform list (TikTok and Instagram).
- Limitations
- Time costs per level are the author's estimates from his own climb, not a measured study.
- As of
- (first published May 22, 2026)
Source: Virlo, AI marketing agents: the 5 levels of agentic marketing, and how to climb each one, November 2026, virlo.ai/resources/agentic-marketing
Want the full list?
Give your marketing agents live TikTok and Instagram data from 23.3M viral videos. Virlo is social listening for TikTok and Instagram Reels: start a 7-day free trial at app.virlo.ai and see real videos from your niche before you pay.
Frequently asked questions
What is agentic marketing?
Marketing work done by AI agents that plan and carry out multi-step tasks, such as research, briefs and scoring, with a person setting goals and approving output.
What are AI marketing agents?
AI systems given a goal, tools and context that run marketing workflows on their own: pulling data, finding patterns, writing briefs and checking them against brand rules.
What are the 5 levels of AI agents in marketing?
Generic prompts, detailed skill docs, skills plus a brand foundation, agents with a live data layer, and autonomous agent fleets with shared memory.
How do I know which level I'm on?
Answer the 4-question self-audit: where ideas come from, your daily keyboard hours, what happens when you take a week off, and what a new junior marketer would do. Your most frequent answer is your level.
What is a brand foundation file?
One document with positioning, voice rules, banned phrases, your dialect and your audience's, customer quotes, ICP segments, competitors, receipts and hard bans, which every skill and agent reads.
Which jump is the hardest?
Level 2 to level 3, because it means writing your own taste down for the first time. The other jumps are mostly mechanical.
What agents does a marketing workflow need?
A common level 4 stack has a research agent, a hook agent, a brief agent and a critic agent.
What does human over the loop mean?
Instead of doing each step yourself (human in the loop), you set the goal and review what the agents produce.
Do AI marketing agents need live data?
Yes. Without a data layer, agents work from stale training data. A source like Virlo gives them current TikTok and Instagram videos, hooks and outlier scores.
Related free resources
- Content automation with Claude: the Monday Drop install kit
- Claude skills for marketing: 50 social media skills for the whole c...
- Claude Code for marketing: a 50-agent social media team in one folder
- Claude code automation: find a viral video format and remake it as ...
- UGC agent: build one in Claude Code that finds outlier creators eve...
- Claude for marketers: the 5 plays a marketing team runs every week
- All Virlo resources
- TikTok trends this week
Run this on your own niche
Give your marketing agents live TikTok and Instagram data from 23.3M viral videos.
7-day free trial, easy to cancel.