The pattern
In March 2026, Andrej Karpathy open-sourced a 630-line script that ran AI experiments while he slept. The agent modified training code, ran a five-minute experiment, checked the result, and looped. He woke up to 700 changes and 20 improvements he'd missed after weeks of manual tuning.
Within ten days, builders had pointed the same loop at financial markets (+22% return), chess engines (expert to grandmaster), and rendering pipelines (53% faster). Nobody coordinated this. The structural conditions were right: a metric, a clock, and a loop.
The pattern has a name now — autoresearch. And it turns out call buying has the structural conditions to run it.
Why call buying has the right conditions
Autoresearch works when three things are true about your domain:
The metric is fast. You need to evaluate each experiment quickly enough to run dozens per day. In Karpathy's case, validation loss evaluated in five minutes. In call buying, CPA evaluates on every call — you know whether a bid change worked within hours, not weeks.
The metric is unambiguous. The number has to be clearly better or clearly worse. Validation loss goes down = better. CPA goes down = better. "Is this email good?" is ambiguous. "Did this bid modifier reduce cost per acquisition?" is not.
The metric is predictive. It has to correlate with what you actually care about. CPA directly measures the cost of acquiring a customer. It's not a proxy — it's the thing itself. This is rare. Most business metrics are proxies. CPA is honest.
Call buying has all three. That's why it compounds.
How Q Optimizer runs the loop
Q Optimizer is an AI agent that observes your campaign performance, proposes experiments, and scores them against your target CPA. You decide what goes live.
The cycle
Observe. The agent reads your campaign results: which geos are converting, which sources are underperforming, which dayparts are overpriced, which bid modifiers are stale.
Propose. The agent proposes a specific change: raise the Florida geo modifier from +15% to +25%, cut the overnight daypart bid by 30%. Each proposal is a discrete, testable hypothesis.
Evaluate. Locked rules, written in code the AI cannot change, score each proposal against your target CPA. The agent can suggest anything. It cannot grade its own work.
Decide. You accept or dismiss the proposal. Accepted changes land in the campaign history, where you can see who changed what and restore any earlier version.
Repeat. Each cycle starts from the results of the last. Housekeeping runs every 15 minutes, and a deeper AI pass runs once each weekday. A per-campaign circuit breaker can stop the agent on any campaign.
Why the gate matters
Karpathy's loop could run unattended because a bad experiment cost five minutes of compute. A bad bid change costs real money on real calls, in a regulated market. So the loop keeps a person at the gate. The agent does the tireless part: reading every geo, source and daypart, every day. You do the part that should stay human: deciding what to spend.
The clock
MMNTM's analysis of the autoresearch pattern identifies the clock, the fixed time constraint, as the mechanism that makes the loop work. Without a fixed clock, experiments are incomparable. A good 3-minute experiment and a good 3-hour experiment can't be ranked on the same leaderboard.
In Karpathy's case, the clock is five minutes (one training run). In call buying, the clock is set by call volume: an experiment needs enough calls on each side of the comparison to tell a real CPA change from noise. Low-volume campaigns need longer windows. High-volume campaigns can decide faster. Some outcomes take far longer than a week to arrive; judging sources by 90-day retention covers that slower check.
Where the loop breaks (and where it doesn't)
The enterprise eval gap — the distance between "the loop exists" and "we have a metric the loop can optimize" — is the reason most agent deployments stall. Most knowledge work doesn't have a loss function. "Is this customer support response good?" can't be scored unambiguously.
Call buying doesn't have this problem. CPA is the loss function. It's fast (evaluates per call), unambiguous (lower is better), and predictive (it directly measures acquisition cost). The loop doesn't break because the score is honest.
Where Q Optimizer can't run the loop: campaign types where conversion data feeds back slowly (30+ day sales cycles), campaigns with too few calls to tell a real change from noise, or situations where the target metric isn't CPA (brand awareness campaigns, for example). In those cases, the clock doesn't generate enough signal per cycle to evaluate experiments.
For everything else, where sales come back the same day or the same week, the loop runs and compounds.
The feedback loop is the moat
The Q Optimizer loop doesn't just improve your bids. It improves the data behind them. Every call you buy and every sale you report updates the conversion rates your target CPA bids use. Better rates mean better bids. Better bids mean better calls. Better calls mean more sales to learn from. The loop feeds itself.
This is the same structural advantage Karpathy identified: the companies running overnight loops aren't announcing it — they're shipping the results at 9am. The gap between buyers who review every geo, source and daypart every day and buyers doing manual weekly bid reviews widens quietly.
Getting started
If you're buying calls on RTB exchanges with CPA as your primary metric, your campaigns have the structural conditions for the loop. The requirements:
- Conversion data flowing: Your dialer or disposition system feeds back sale/no-sale outcomes within 24 hours
- Enough volume: Enough calls per campaign that a real CPA change stands out from noise
- Clear CPA target: A number that's unambiguously better when it goes down
- Patience for week 1: The loop needs enough calls to tell signal from noise before its proposals mean much
The full platform overview covers the broader system. Q Optimizer is the piece that makes the system compound.