Frontier system · Marketing · 6 stages · updated 2026-10-23

Self-improving ad creative loop

Performance data in; hundreds of on-brand, limit-checked variants out, rendered into creative by a Figma plugin. A memory log of past hypotheses means each round builds on the last.

1Is this my problem?

Ad testing is limited by how many variants one marketer can write and build, so tests are small and slow.

You'll recognise this if…

A handful of ad variants per month.
Ad fatigue shows before new creative is ready.
Learnings from past tests live in someone's head.

Cost of doing nothingSpend keeps flowing to tired creative while competitors test more.

Companies affectedCompanies spending $10k+/month on Google or Meta ads with a 1-3 person marketing team, including SaaS and DTC brands. Anthropic's growth team runs this with one person.

2Is this solution worth it?

Reported results

Anthropic growth marketing: ad copy went from 30 minutes per ad to 30 seconds (source)

Quoted from the linked source. Results vary by company and setup.

Your ROI, with the assumptions shown

Prefilled with our illustrative assumptions for a typical team (8 h/week of manual work at $60/h, 60% automated). Change them to your numbers.

Build cost (estimate)$8k–$25k

Time to first version (estimate)3-6 weeks

AI agent stages2 of 6

Human checkpoints1

Limits and when not to build it

Not worth it under a few thousand dollars a month in spend, where tests never reach significance.

Guardrails it needs

Character limits validated in code.
Manual review before upload.
Claims checked against product docs.

3How does it work?

The agency builds you one internal tool, Creative Lab, that runs the whole process. Watch it work on a sample case:

Creative LabIllustrative demo with sample data

Pick a sample on the left and press Run example to watch Creative Lab work through a case step by step.

Your tool connects to

DocsAdsAI model

Most of it is ordinary automation. AI agents handle 2 of 6 steps, and people approve at 1 checkpoint.

Under the hood: stages, vendors and the learning loop

AI agent Automation Human checkpoint

1Pull performance
Ads CSV, or Meta Ads data through an MCP server.
Google AdsMeta Ads MCP
Automation
2Pick what to test
Find underperformers and write hypotheses.
Claude Code
AI agentDecides: Pick what to test
3Generate copy
Two sub-agents: headlines (30 characters) and descriptions (90), checked against brand Agent Skills.
Claude CodeAgent Skills
AI agentDecides: Write variants
4Render creative
Figma plugin swaps copy into templates, up to 100 variants at once.
Figma
Automation
5Review
Brand, legal and layout check.
Marketer
Human
6Launch
Upload CSV and assets.
Google AdsMeta
Automation

Learning loop

A memory log of hypotheses and results is loaded into context next round, so the system learns which angles work instead of starting fresh.

Example run

Illustrative sample data showing what moves between stages. Not taken from the sources.

1
Pull performance · inGoogle Ads CSV, 60 ads
outCTR and CPA per ad
2
Pick what to test · inPerformance
out9 underperformers; hypothesis: outcome-led beats feature-led
3
Generate copy · inHypothesis
out30 headlines (all <=30 chars), 12 descriptions (<=90)
4
Render creative · inCopy
out64 creative variants in Figma
5
Review · inVariants
out52 approved
6
Launch · inCSV
outUploaded

Vendors commonly used

Google AdsMeta Ads MCPClaude CodeAgent SkillsFigmaMarketerMeta

Where AI decides and where people do

AI agents decide

Pick what to test: Pick what to test
Generate copy: Write variants

Human checkpoints

Review: Brand, legal and layout check.

Maturity ladder

What each level adds. Most teams start at level 1 or 2 and climb. Highlighted: where this build sits.

Level 1Basic automation
Marketers write ad copy by hand in the ads UI.
Level 2AI-assisted
An LLM drafts headlines that are pasted in.
Level 3Autonomous agent
Sub-agents generate limit-checked variants, a plugin renders creative, a person reviews.
Level 4Self-learning system
A hypothesis log feeds each new round so the system learns which angles win.

Building blocks

Start with one of these smaller workflows and add stages.

Ad copy and creative variant agentAnthropic: ad copy work went from about 30 minutes per ad to 30 seconds; the Figma plugin generates up to 100 variations per batch

An agent reads a CSV of existing ads with performance, finds weak ones, and generates replacements that respect character limits (30 for headlines, 90 for descriptions) using two sub-agents.

MarketingFrontier~4–10 h/week$2k–$8k Coding agent as a marketing team (skills library)

Load an open-source library of marketing "skills" (copywriting, SEO audits, CRO, pricing pages, email) into a coding agent such as Claude Code, Codex or Cursor, and run them against your own site and data..

MarketingAdvanced~3–8 h/week$1k–$5k

Seen in the wild

Links checked 2026-10-23. The pipeline combines patterns from these public builds; we are not affiliated with these companies.

4Who can build it?

Matched on the services, platforms and industries agencies list on their own sites, and on their published projects.