Part 6: The AI Agents
Part 6 of 7

The 6 AI Agents

Deep dive into each agent — what it does, how it works, and what it looks like in action with full interactive demos.

⏱ ~10 min read
🎓 Intermediate Level
🤖 6 Agents

Each of these agents handles one complete process in your media buying workflow. They can work independently or be chained together into automated workflows (covered in Part 7).

Agent 01

The Creative Agent

Generates ad copy, headline variations, and creative concepts in seconds — not hours.

The Creative Agent handles the most time-consuming part of media buying: coming up with ad variations. Instead of spending hours writing copy and brainstorming angles, you describe the product and get dozens of variations back.

Capabilities

  • Generates primary text (ad body) in multiple creative angles
  • Creates headline and description variations for A/B testing
  • Adapts tone: cold prospecting vs warm retargeting vs win-back
  • Suggests angles: pain points, benefits, social proof, urgency, FOMO, authority
  • Describes permitted public creative patterns without inferring spend, performance, intent, or copying protected expression
  • Formats for different placements: Feed, Stories, Reels
creative agent — generating ad copy
You: Generate ad copy for premium leather bags, women 25-40 Egypt, 2500-4000 EGP
⚙ Generating creative variations across 4 angles...
Angle 1: Quality & Craftsmanship
Primary: "[Verified material]. [Substantiated production detail]."
Headline: "[Approved durability benefit—only if evidence supports it]"
Angle 2: Social Proof
Primary: "[Verified customer count] customers; [current rating] from [verified review count], sourced [date]."
Headline: "[Approved social-proof line]"
Angle 3: Urgency / Scarcity
Primary: "[Live available-to-promise inventory—include only while true]."
Headline: "[No fabricated urgency or scarcity]"
Angle 4: Pain Point
Primary: "[Approved customer problem] + [substantiated product benefit]."
Headline: "Stop Replacing — Start Investing"
✓ 12 total variations generated (4 angles x 3 formats)
💡 Pro Tip

Use a versioned brand brief, approved claims library, product source data, and reviewed examples. Keep each client's data tenant-isolated; do not silently train on, memorize, or reuse client inputs. Every factual, comparative, scarcity, testimonial, and performance claim needs source evidence and human approval.

Agent 02

The Campaign Launcher

Creates complete campaigns on Meta — without touching Ads Manager.

The Campaign Launcher handles the full campaign creation process via the Meta Marketing API: choosing objectives, setting budgets, configuring targeting, creating ad sets, and linking creatives.

The 5 Campaign Commands

CommandWhat It Does
/create-campaignCreates campaign with objective + budget (CBO or ABO)
/setup-adsetConfigures targeting, placement, optimization goal
/create-adCreates ad creative and links to ad set
/check-performancePulls metrics and flags anomalies
/scale-campaignDrafts a capped canary request; cannot authorize spend
campaign launcher — building a full campaign
$ /create-campaign sales 800 EGP/day Egypt women 25-40 fashion
⚙ Creating via Meta Marketing API...
✓ Campaign: "Sales — Fashion — Apr 2026"
Objective: OUTCOME_SALES | Budget: 800 EGP/day (CBO) | Status: PAUSED
✓ Ad Set: "Fashion — Women 25-40 — Egypt"
Targeting: Egypt, Women 25-40, Interest: Fashion | Optimization: Purchase
✓ Ready for review in Ads Manager
⚠ Campaign is PAUSED — review and activate when ready
🛡 Safety First

A prompt cannot guarantee safety. A tenant-scoped executor rejects ACTIVE creation, unapproved accounts, spend above the cap, stale or mismatched approvals, and duplicate requests. It hard-codes PAUSED, requires an idempotency key and exact signed change-set, logs the write, and stores rollback. Activation is a separate privileged approval.

Agent 03

The Budget Optimizer

Use evidence to draft account-specific budget proposals; data and model output do not authorize a spend change.

The Budget Optimizer can summarize reported ROAS alongside contribution margin, attribution and conversion lag, incrementality, sample size, objective, cash flow, and account history. It proposes a reversible canary; a human approves the exact diff and a service enforces the cap, stop condition, idempotency, and rollback.

Illustrative Display — Not Universal Policy

ROASCategoryAction
Above example targetInvestigate capacityConsider a capped canary
Near example targetReviewCheck lag, margin, and sample size
Below example targetInvestigateCheck causes and measurement
Below stop criterionPropose holdHuman-approved, reversible action only
budget optimizer — performance report
$ /optimize-budgets last 7 days
Campaign Performance Report — Last 7 Days
CampaignSpendROASAction
DPA Retargeting5,2004.2xINVESTIGATE
Summer Collection3,8003.1xCANARY?
UGC Testing1,2001.8xREVIEW
Brand Awareness8000.3xHOLD?
💰 Draft only: evaluate a capped, reversible reallocation canary
📈 Impact is uncertain; show evidence, assumptions, holdout, and rollback
⚠ Issue structured change request for exact owner approval
💡 Real Talk

Reported ROAS alone cannot establish profit or incremental lift. Calibrate criteria per account, inspect uncertainty and lag, and prefer small canaries over universal scale/pause rules.

Agent 04

The Performance Monitor

Catches problems BEFORE they drain your budget — not after.

The Performance Monitor can evaluate scheduled, account-calibrated checks and flag observations for investigation. Detection speed depends on data freshness, sample size, and conversion lag.

What It Watches

📈
CPM Deviation
Compare with a seasonality-aware baseline and minimum sample. Auction, mix, delivery, attribution, lag, or site changes are hypotheses.
🔄
Frequency Change
Use an account-specific baseline by audience and objective. Frequency alone does not prove fatigue.
📉
CTR Deviation
Check placements, mix, delivery, sample size, and creative before proposing a reversible test.
💰
Outcome Review
Reported ROAS requires margin, incrementality, attribution-lag, sample-size, and objective context; it never triggers an automatic write.
performance monitor — anomaly detection
⚠ ANOMALY DETECTED — "Summer Collection CBO"
Issue: CPM spiked 65 → 112 EGP (+72%) in 24 hours
Impact: CPA increased 180 → 310 EGP
Hypothesis: frequency may indicate saturation; corroborate with breakdowns, changes, lag, and delivery
💡 Recommended Actions:
1. Refresh creatives on saturated ad set
2. Expand lookalike from 1% to 3%
3. Add new interest targeting layer
Agent 05

The Report Generator

Generate a tenant-isolated draft with sources, definitions, freshness, and uncertainty for human review.

The Report Generator may read only sources authorized for that tenant and report purpose. It must not join, compare, or reuse data across clients. Treat the output as a draft: a named reviewer checks access scope, metric definitions, attribution, claims, and recipients before export or send.

Report Types

Daily
Morning Briefing
Quick summary: spend, revenue, ROAS, top/bottom performers, any alerts. Takes 30 seconds to generate.
Weekly
Performance Analysis
Detailed analysis with trends, week-over-week comparisons, winner/loser identification, and recommended actions.
Monthly
Strategic Review
Comprehensive report with channel performance, audience insights, creative analysis, budget recommendations, and next month strategy.
report generator — daily briefing
$ /daily-report
⚙ Pulling data from Meta Ads API...
⚙ Analyzing 6 campaigns, 18 ad sets...
Daily Performance — April 3, 2026
Total Spend: 4,200 EGP | Revenue: 14,700 EGP | ROAS: 3.5x
▲ +12% vs yesterday | +8% vs last week avg
Top Performer: "DPA Retargeting" — 4.8x ROAS
Needs Attention: "UGC Test B" — frequency at 3.1
✓ Tenant-scoped report draft created with source + freshness notes
✓ No export or send until reviewer confirms recipients and content
Agent 06

The Scaling Engine

Ranks candidates and drafts controlled canary proposals; it does not scale or retire campaigns automatically.

The Scaling Engine evaluates candidates using account-specific readiness criteria and uncertainty. It can propose a PAUSED test or capped budget canary, but a tenant-scoped service requires an exact approval, enforces spend and idempotency limits, monitors stop conditions, and can roll back.

Scaling Strategies

StrategyHow It WorksWhen to Use
VerticalPropose an account-capped budget canaryIncrementality, margin, lag, and capacity pass
HorizontalDraft a PAUSED audience testEvidence supports a distinct hypothesis
CreativeDraft substantiated creative variantsRights, claims, and eval criteria pass
RotationPropose a reversible rotation testAccount-specific fatigue evidence is corroborated

The Winner Identification Loop

This is the most powerful pattern — a continuous cycle of finding, analyzing, replicating, and scaling winners:

scaling engine — the winner loop
Step 1: Identify winners (high CTR + high ROAS + low frequency)
→ "Summer Lifestyle Video" — CTR 2.8%, ROAS 4.2x
Step 2: Extract the winning DNA
→ Hook: lifestyle context | Format: video | CTA: Shop Now | Duration: 15s
Step 3: Generate new variations with same DNA
→ 6 new creatives: same style, different products/angles
Step 4: Create PAUSED draft through controlled executor
→ Allowlist, cap, idempotency, approval, audit, and rollback enforced
Step 5: Evaluate canary, uncertainty, and rollback criteria
→ 3 of 6 candidates; owner decides the next exact change
🔄 Loop back to Step 1 — continuous improvement
Illustrative metrics only; no ROAS progression or budget path is guaranteed.
🔬 Why This Loop Matters

A recurring evaluation can keep a queue of candidates, but each test and spend change needs evidence, an account-specific readiness gate, human approval, deterministic controls, observation, and rollback. Continuous automation does not guarantee momentum or performance.

6 Agents. One System.

In Part 7, we tie everything together — daily routines, chained workflows, and how to go from here to running a fully AI-powered media buying operation.

Built by @itsmazinzaki — AVAMARTECH