Andrej Karpathy released autoresearch in March 2026. The idea is deceptively simple: give an AI agent a task, a metric, and the ability to change things, then let it run experiments while you sleep.
In Karpathy’s case, the agent optimized neural network training code. In one overnight run, it completed 126 experiments. Over two days, it processed 700 autonomous changes, found 20 improvements that stacked together, and achieved an 11% efficiency gain. No human intervention. The agent formed hypotheses, tested them, kept what worked, discarded what didn’t, and repeated.
The pattern itself has nothing to do with neural networks. It works anywhere you have three things: a number that goes up or down, something you can change, and an API to connect them. That describes most of the business processes companies spend millions trying to optimize.
The Pattern in One Paragraph
Start with a baseline. Measure it. Generate a hypothesis for improvement. Change one thing. Measure again. If it’s better, keep it and log why. If it’s worse, discard it and log that too. Generate a new hypothesis informed by everything learned so far. Repeat indefinitely.
This is what good marketing teams do manually over weeks. The autoresearch pattern does it in hours, without human bottlenecks between each step.
Where It Creates Business Value
Cold Email Outreach
The metric: Reply rate. What changes: Email copy – subject line, opening line, offer, length, call to action. Business value: A 1% increase in reply rate on 10,000 monthly sends is 100 more conversations. At a 10% close rate and $5,000 average deal size, that’s $50,000 in additional pipeline per month.
How it works: The agent writes a variant email, sends it to a sample via the email platform’s API (Instantly, Apollo, Lemlist), waits 48 hours, pulls reply rate data, and scores the variant against the baseline. It logs what it learned – “emails under 75 words with a specific question in the subject line outperform by 23%” – and generates the next variant using accumulated learnings.
A marketing team tests one variant per week. The agent tests 3-5 per day.
Landing Page Conversion
The metric: Sign-up or purchase conversion rate. What changes: Headline, hero text, CTA button copy, social proof placement, form length. Business value: A landing page converting at 3% instead of 2% means 50% more leads from the same traffic. If you’re spending $10,000/month on ads driving traffic to that page, the same budget now produces 50% more customers.
How it works: The agent edits page elements via the hosting API (Webflow, WordPress, or a custom CMS), deploys the variant, waits for statistically significant traffic, compares conversion rate to baseline, and keeps the winner. Over a week, it can test 10-20 variations – more than most companies test in a quarter.
Ad Creative Optimization
The metric: Cost per acquisition (CPA) or click-through rate (CTR). What changes: Ad copy, headline, image selection, call to action. Business value: Dropping CPA from $50 to $40 on a $20,000/month ad budget means 100 more customers per month for the same spend. Or the same customers for $4,000 less.
How it works: The agent calls the ad platform API (Meta, Google Ads), creates new ad variants, monitors performance metrics, pauses underperformers, and scales winners. The feedback loop is fast because ad platforms generate data within hours.
Customer Support Scripts
The metric: Customer satisfaction (CSAT) score or resolution rate. What changes: Response templates, escalation triggers, tone, suggested solutions. Business value: A 5-point CSAT improvement reduces churn. For a SaaS company with 1,000 customers at $200/month, reducing monthly churn from 3% to 2.5% saves $12,000/year – and that compounds.
How it works: The agent modifies the response templates that support agents (human or AI) use. It tracks downstream CSAT scores per template variant. Templates that produce higher satisfaction replace the baseline. Over weeks, the “master script” evolves toward what actually resolves issues.
Product Descriptions (E-commerce)
The metric: Sales per product page view. What changes: Product description copy, feature emphasis, benefit framing. Business value: An e-commerce store with 500 products averaging 100 views/day and a 2% conversion rate generates 1,000 sales/day. Improving conversion to 2.3% adds 150 sales/day – $54,750/year at $1 profit per unit. From better words on a page.
How it works: The agent updates product descriptions via the store’s API (Shopify, WooCommerce, Amazon Seller Central), monitors sales data, and iterates. Products with the most traffic get optimized first for maximum impact.
Pricing Page Optimization
The metric: Conversion to paid plan, or revenue per visitor. What changes: Price points, plan names, feature bundling, CTA copy, layout. Business value: The pricing page is often the highest-leverage page on a SaaS website. A 10% improvement in conversion rate at that step directly increases revenue by 10%. No additional traffic, no additional ad spend.
How it works: The agent edits pricing page elements, monitors sign-up and upgrade data, and moves the page design toward whatever configuration generates the most revenue per visitor. It might discover that renaming “Pro” to “Growth” and moving the enterprise CTA above the fold increases upgrades by 8%.
Why This Is Different From A/B Testing
Traditional A/B testing has one bottleneck: humans.
A typical A/B test cycle takes five weeks. A human writes the hypothesis. A human designs the variant. A human sets up the test. A human waits for data. A human analyzes the results. A human decides what to test next. Five weeks for one data point.
The autoresearch pattern removes the human from every step except defining the metric and reviewing results. The agent generates hypotheses based on accumulated learnings, not guesswork. It creates variants, deploys them, collects data, evaluates results, and generates the next hypothesis – all without waiting for someone to schedule a meeting about it.
The result: 10-50x more experiments in the same time period, with each experiment informed by every previous one. That’s not a marginal improvement. It’s a structural advantage.
What It Requires
Three non-negotiable ingredients:
1. A clear, objective metric. Reply rate, conversion rate, revenue, CSAT, cost per acquisition. The metric must be a number that goes up or down. “Brand awareness” or “engagement” don’t work unless you can reduce them to a measurable signal.
2. A controllable input. The agent must be able to change something: email copy, page content, ad creative, pricing, templates. If changes require a designer to manually update an image in Figma, the loop breaks.
3. A programmable surface. An API or scripted interface to both read metrics and apply changes. Email platforms, ad platforms, CMS systems, analytics APIs. If the only way to change a landing page is to edit HTML by hand and FTP it to a server, the pattern doesn’t apply.
What Doesn’t Work
Slow feedback. If each experiment takes a month to produce data (enterprise sales cycles, annual contract renewals), the loop is too slow to be useful. The pattern thrives on fast cycles – minutes to days, not weeks to months.
Fuzzy metrics. If you can’t define “better” as a number, the agent can’t optimize. “This email feels warmer” is not a metric. “Reply rate increased from 4.2% to 5.1%” is.
No API access. If the system you want to optimize can’t be changed programmatically, you need humans in the loop, which defeats the speed advantage.
The Cost of Not Doing This
Every business is already running this optimization loop manually. Marketing teams write emails, wait for data, adjust, and try again. Product teams tweak pricing pages based on quarterly reviews. Support teams update scripts when CSAT drops.
The autoresearch pattern doesn’t do anything fundamentally new. It does the same thing 10-50x faster, with no gaps between experiments, and with a memory that accumulates learnings across every iteration instead of losing context when someone goes on vacation.
The companies that adopt this pattern will optimize their customer-facing surfaces continuously and automatically. The companies that don’t will optimize quarterly, manually, and partially. Over 12 months, the compounding difference in conversion rates, reply rates, and acquisition costs will be substantial.
The barrier to entry is not technology. It’s willingness to define a metric, connect an API, and let the loop run.
Sources: