Running every 15 minutes

Autonomous
optimization

Q is an AI agent that observes your campaign performance, proposes bid experiments, evaluates results against locked scoring functions, and applies winning changes — autonomously, every 15 minutes, 24/7.

Get started See how it works

The idea

“What if an AI agent could run the same optimization loop a senior media buyer runs — but do it every 15 minutes instead of once a week?”

Inspired by Karpathy's autoresearch pattern. Q observes campaign performance, plans bid experiments, evaluates results against locked scoring functions, and applies winning changes — autonomously, 24/7, with guardrails the agent can't override.

The optimization loop

Every 15 minutes, Q runs a complete observe → plan → experiment → evaluate → decide cycle for each active campaign.

Observe

Q reads segment performance data — calls, conversions, CPA by geo, hour, source, and subID. It reviews the last 50 experiments to learn what worked and what didn't.

Reads last 50 experiments + live segment data

Plan

Based on observed patterns, Q identifies high-opportunity segments. "Florida converts at 73% vs 31% national — the geo modifier should be higher."

AI reasoning over performance patterns

Experiment

Q proposes a bid modifier change — e.g., increase FL geo modifier by +20%. The system validates against 5 hard constraints before applying.

e.g. FL geo modifier +20%

Evaluate

After accumulating enough calls (50+ minimum), Q evaluates the experiment. The locked evaluator compares treatment vs. baseline on proxy score and CPA.

Locked scoring — agent cannot modify

Decide

Q recommends commit or revert. The system checks the locked evaluator independently — if the evaluator disagrees, it rejects the agent's recommendation. Safety first.

Commit winner or revert to baseline

What Q actually does

Real experiment log from a single campaign. Q ran 47 experiments over 2 weeks — 31 committed, 12 reverted, 4 expired.

Medicare AEP — Nationalautonomous

Last cycle: 2 min ago

committedgeo_state:FL+20%312 calls

FL converts at 73% vs 31% national. Increasing modifier to capture more volume.

$47.80 → $42.10

-12% CPA

runninghour_of_day:10-14+15%89 calls

Peak conversion window 10am-2pm ET. 62% of conversions occur here.

$44.20 → $41.50

-6% CPA

revertedsource:mediaalpha-25%156 calls

Source quality degraded. Reducing modifier to limit exposure. Evaluator confirmed no improvement.

$38.90 → $41.20

+6% CPA

runninggeo_state:TX+10%34 calls

TX shows improving conversion trend. Testing moderate bid increase.

$36.40 → —

committedhour_of_day:22-06-30%201 calls

Overnight calls convert poorly. Reducing bids to preserve budget for peak hours.

$62.10 → $58.40

-6% CPA

Marketing Autopilot

Most experiments fail — that's by design. Q runs dozens per campaign, reverts what doesn't work, and compounds the improvements that do. Your CPA drops while you sleep.

Q OPTIMIZER · MEDICARE AEP — NATIONAL67 EXPERIMENTS · 0 KEPT

$58.20

CURRENT CPA

-0.0%

CPA DROP

KEPT

Safety architecture

The locked evaluator

Q proposes changes. But a separate scoring function — invisible to the agent, immutable by the agent — validates every decision before it touches your campaigns.

The creative function (plan, experiment) is separated from the evaluation function (score, validate). The agent can't game what it can't see.

Hard limits (agent cannot bypass)

Scoring thresholds — proxy score and CPA limits

Modifier delta limits — max ±30% per experiment

Bid floor — effective bid never drops below minimum

Concurrent experiment cap — max 3 per campaign

Minimum call threshold — 50+ calls before evaluation

Dimension allowlist — only approved segment types

What the agent decides (creative judgment)

Which segments to experiment on

How much to adjust modifiers (within limits)

When to commit or revert experiments

Pattern recognition across experiment history

A fully automated marketing team

Q doesn't just optimize silently. It keeps you in the loop — Slack messages, email digests, or SMS alerts with what changed, why, and what it plans to do next.

Most teams go from shadow to autonomous within a week.

Shadow

Advisor

Q runs the full optimization loop but doesn't apply changes. See what Q would have done — build confidence before going autonomous.

Full experiment proposals logged
No modifiers changed
Review "what would have happened" in dashboard
Most teams go autonomous within a week

Supervised

Copilot

Q proposes experiments, but you approve before they run. See everything Q is thinking — same AI, your final call.

Proposals require your approval
Commit/revert recommendations visible
Override any decision
Available if your compliance requires it

Autonomous

Autopilot

Q runs independently, 24/7. The locked evaluator validates every decision. Changes are applied automatically. Q messages you via Slack, email, or SMS.

Locked evaluator validates every decision
Auto-revert on timeout (14 days)
Campaign-level kill switch always available
Performance updates via Slack, email, or SMS

Autoresearch, adapted for performance marketing

Karpathy's autoresearch pattern lets an AI agent iteratively improve code. We adapted it for bid optimization — same architecture, with guardrails built for real ad spend.

Concept	Autoresearch	TapQuality
Playbook	Locked operational manual	Campaign optimization policy — your rules, Q's boundaries
Evaluator	Locked scoring function	Locked scoring — proxy score + CPA thresholds the agent can't see
Experiment	Mutable code changes	Bid modifier adjustments — the levers Q is allowed to pull
History	Append-only results log	Full experiment audit trail — every proposal, result, and decision
Commit	Keep what worked	Winner applied — modifier stays, baseline updated
Revert	Undo what didn't	Loser rolled back — modifier restored to pre-experiment value

The numbers

experiments/day

Per campaign at 15-min cycles

14%

avg CPA reduction

Within first 2 weeks

< 60s

cycle time

Observe → decide in under a minute

manual interventions

In autonomous mode