An agent can now run the whole revenue loop in Adapty

TL;DR:
- An agent can now run the revenue loop in Adapty from Claude Code, Codex or the Claude app: audit, hypothesis, products and prices, flow, segment, A/B test, result, next hypothesis.
- The agent works from data on 30,000+ apps. In those apps, localization, trial and plan-length tests raised LTV about 60% of the time. Visual and copy tests raised it 35% of the time.
- You approve every change. Each hypothesis is saved with its reasoning to a list your team reviews in the dashboard.
- The loop runs on a schedule too: a weekly health check and checks on live tests, with the report sent to Slack or email.
I have run Adapty's growth products since the first Autopilot, and teams have asked for the same thing the whole time: tell me what to do next and where the most money is, then do it for me with one button. This month a client sent me the same idea as a product we should build. Until this year we could only do the first half. This year coding agents made the second half possible, and we shipped the first version in October.
An agent can now run the whole revenue loop for a subscription app inside Adapty. The revenue loop is the cycle an app repeats to grow revenue: audit, form a hypothesis, build it, test it, read the result, pick the next test. A large share of new flows on Adapty are built by an agent today.
What does a closed revenue loop look like?
An agent takes your app from audit to the next hypothesis inside one Adapty account. You install the Adapty plugin in Claude Code, Codex or another coding agent, or connect the Adapty MCP server in the Claude app, and ask a question in plain language. The Growth Advisor skill reads your analytics, prices, placements, flows and past A/B tests, then builds what the next test needs.
- Audit. The agent reads your revenue, conversion, renewals and refunds and grades each against the market.
- Hypotheses. It ranks them by the dollars a month each one puts at stake and saves the ones you keep to your hypothesis list.
- Product and prices. It creates a copy of the product in the store and sets prices by country.
- Flow. It builds a copy of the flow with one change.
- Segment. It creates the audience for the test.
- A/B test. It sets up the A/B test and shows who sees what.
- Result. It reads revenue per variant.
- Next. It suggests the next hypothesis for that placement.
Every step that changes your account waits for your approval.
On the sample photo app from our docs, the audit found an annual plan at $29.99, cheaper than 90% of US photo apps. The agent proposed a test at $33.99 for US users. After the owner approved each step, it created the product, the flow copy and the test.
A different app gets a different first answer, because the agent starts from your own data: where the app loses money today and where it can earn more now. For one app that is a price, for another a trial, a plan ladder or a paywall shown in onboarding. When an app has too little history, the agent proposes what most often works for similar apps and labels its numbers as estimates.
Teams that run their own experiments can put their own hypotheses through the loop. Tell the agent what you want to test, and it grades the idea against the same evidence, builds the variant, sets up the A/B test and reads the result by the same rules.
Why did we build an agent and not a better dashboard?
Our first Autopilot gave recommendations, and teams launched few of them. It benchmarked an app against the market and returned a testing plan. Where teams launched the plan, it paid: AppDevLabs grew MRR 27.5% in three months. Customer interviews explained the rest. Starting a test took a click, and preparing the product, the paywall and the segment took the week.
A better recommendation screen would not fix that. The work had to go away. Three things made that possible. Coding agents learned to carry a task across many steps, and growth teams began to keep one open all day. The Adapty CLI and MCP server gave an agent the same access to an account that a person has in the dashboard. The Growth Advisor skill gave it the judgment our Professional Services team uses with clients.
We shipped the first Autopilot at the end of last year. The first version of Growth Advisor followed in October, and the loop closed in November.
What does the agent know that your own data can't tell it?
It knows how the rest of the market converts, prices and renews, and which tests won there. Your own events tell an agent that your paywall converts 3.1% of views. They don't tell it that the US median is 2.75% and the European median is 1.38%.
The numbers come from Adapty's State of Paywalls 2026, which covers 78,000+ paywalls from 30,000+ apps, from State of Onboarding 2026, and from our price index for 60+ countries.
The same report counts how often each kind of A/B test raised LTV.
So the agent puts an offer change ahead of a redesign. It starts with onboarding and the first paywall, because 93% of first purchases happen within 24 hours of install. It suggests a price raise only in Tier-1 markets, where a raise tends to lift ARPU by 15% to 25% and has lost elsewhere.
Onboarding gets the same treatment. The agent sets each screen of your flow against State of Onboarding 2026, with medians by region, category and flow length, and turns each gap into a hypothesis. In that report the first screen loses 10.2% of sessions, more than five times the loss on any screen after it.
Each number in a reply names its source, and an estimate is labeled as an estimate.
Benchmarks are half of it. The other half is how the agent reads your data. Ask a top model without the skill whether your paywall is the problem, and it tends to bend the answer to your question. Growth Advisor checks the other causes first: it splits a falling rate by channel before it blames the product. It judges a change against fixed thresholds, and it calls a rate with fewer than about 20 purchases behind it too few to tell. A 150% jump from a small base is noise.
Why does a closed loop earn more?
Revenue grows from many small tests, and the tool that runs them is a mechanism. Pregnancy+ by Philips went from one paywall test a quarter to three or more a month when a no-code builder took engineers out of the loop. Weekly subscription revenue grew 300% in 12 months. The builder removed the release. The agent removes the setup: the product copy, the flow copy, the segment, the test.
Speed alone would produce more bad tests. The agent builds one change per hypothesis, so a result names its cause. When a test ends it lines up the next one, so your main placement always has a test running.
Count the tests that ran on your main placement last year. Among apps with 50,000 to 500,000 paywall views a month, the top quarter ran 9 or more. Above 500,000 views, the top quarter ran 21 or more.
Can the agent check your app on a schedule and report to Slack, email or Teams?
Yes. After a health check the agent offers to repeat it every Monday, and once an A/B test is live it offers to check the test twice a week. Your own agent runs each check on its scheduler and sends the report through any connector it holds: Slack, email, Teams, Notion, Linear or another one. A scheduled run only reads your data until you answer it. The docs describe how recurring checks work.
This is why Growth Advisor runs in your own agent, and why we did not build a chat for it into the dashboard. Your agent already has your connectors, your scheduler and your other tools, so the list of places a report can go is yours to extend. A report to a messenger we have never heard of needs no integration from us. The skill also fits the setup you have, whether that is Claude Code, Codex or another coding agent.
Who decides what the agent runs?
You do, and your team does it in one place. Each hypothesis you keep goes to a list in AI Growth Advisor in the Adapty dashboard with the reasoning behind it. Anyone on your team can open the list and read why the agent proposed a test before it runs. In a large company the person who decides is often not the one at the terminal.
The agent changes nothing in your account on its own. Reads are read-only, and each write waits for you to confirm it.
Three refusals are built in:
- It won't start an A/B test before you see who gets which variant.
- It won't name a winner before each variant has about 200 purchases.
- It won't test a price on your live product. A bad price test hurts subscribers you have, so the test runs on a copy and they keep their price.
How do you try it on your app today?
Install the plugin with one command and ask for an audit.
Restart your agent, then ask the question from the top of this post:
No terminal? Add the Adapty MCP server in the Claude app and sign in to your Adapty account. There is nothing to install, and the same question works there.
No Adapty account yet? Send a store link:
The reply is a scorecard against the market and a list of hypotheses in the order to test them. Pick the first one, and the agent sets up the test.



