A growth dashboard tells you what happened and leaves the next step to you. Most founders who built their product with a coding agent never take that step, because the next step is a judgement call nobody taught them to make. A growth agent should make that call, ship it as one small pull request, and then read what it did.
That is the bet behind the agent we build. This post is about why we chose a pull request over a chart, how the agent picks the one change, and the rules that keep it honest.
The dashboard leaves the hard part to you
Picture the founder we build for. They shipped a product with Claude Code or Codex in a few weekends. It works. Traffic trickles in, a few people sign up, nobody pays. They open their analytics and see a funnel with a big drop between the signup page and an actual signup.
Now what? The chart is correct and it does not help. Is the drop the form, the copy, the button order, or the people arriving at all? Which of those should they fix first, and how would they know it worked? A dashboard answers "what" and stops. The founder who could answer "so what do I do on Monday" is a growth person, and they cost a salary or a retainer.
The coding agent changed one side of this. Building got cheap. Boris Cherny, who created Claude Code at Anthropic, ships 20 to 30 PRs a day, all of it AI-written. At Anthropic, code output per engineer grew 200%. The other side did not move. Deciding what to build for growth is still the slow part, and a dashboard does not make it faster. It gives you more to read.
So we asked a simple question. If the agent already lives in the founder's repo and can open a pull request, why should it hand back a chart instead of a change?
One change a day, small enough to review
The daily pull request is an old habit for engineers. Philip Rideout, a professional developer for 23 years, wrote that the daily PR forces pull requests to be small and focused, and that it is more of an attitude than a strict metric. He also wrote that one of the most demoralizing things a developer can feel is being stuck.
Founders feel that second part about growth every day. A long list of things to try is a way to stay stuck. One change, already written, waiting for review, is a way out.
Small growth changes can matter. In one account, a PM changed two words on a call to action and got a 14% conversion lift. Most changes do less than that, and some do nothing. That is fine as long as each one is small enough to review in a few minutes and isolated enough to read on its own.
A pull request also gives the founder the right kind of control. They do not have to agree with a strategy deck. They look at a diff. The page copy changed from this to this, the signup form now leads with GitHub, the sitemap now lists the new page. They merge it or they do not. Planning moves to where the work already is. Rippling's CPO posted that planning decks have been killed and planning now happens in git, and for a solo founder that is even more true.
Not every growth move is code. A directory listing, a launch post or an email to someone who asked for what you sell happens in a browser or an inbox. Our agent fills those in and the founder presses submit or send. But the default shape is the same: one finished piece of work, ready for a yes.
How the agent picks the one move
The pick is the part that matters, and the part a dashboard skips. Here is how ours does it, in order.
It reads the numbers first. Before it looks at any move, the agent writes down what holds the product back this week, from the product's own numbers: traffic by channel against the month before, each step of the funnel, and, once the product reports its users, which channel's signups actually go on to use it and pay. A channel's clicks are not its customers. That diagnosis becomes the first line the founder reads.
It only chooses from moves with real work behind them. A rules layer finds every move that is open for this product: the checks still failing on the live site, the directories it is not listed in yet, the page worth writing, the test waiting to be read. The model cannot invent a move. It ranks what the rules found.
It weighs evidence in a fixed order. The founder's own results come first: what this kind of move did for them before. Then people who asked for it in public. Then what products a step ahead of them do. Then a library of plays traced to other companies. A move with no outside play behind it is not ranked lower for that. And above all of it sits what the founder typed to the agent in their own words, which it reads as their direction.
Every number in its reasoning has to be real. When the agent writes how many visitors leave the signup page, that count must be in the facts it was given for that move, or the sentence is dropped and the plain version is kept. This sounds small. It is the difference between an agent you can trust and one that sounds confident.
It offers ten, not one, and the list changes. We tried offering a single move. Founders wanted to see what else was there. So it offers ten moves, with the first as its pick and its reason for each. A kind of move the founder was shown on three days in two weeks and never took rests a week, so the choices do not go stale.
The founder can ignore the pick. They often do, and the agent keeps count of what came of those calls next to its own. That record matters more than any single choice.
Why it waits instead of stacking changes
The easiest way to learn nothing is to change several things on the same page in the same week. If signups go up, which one did it? If they go down, which one do you roll back?
So the agent holds changes back on purpose. A home page change is read on its funnel step 28 days each side, and any other change to the same path waits until that read is done. The founder sees those held moves in the daily brief with the date they open up, so not changing something is visible as a decision, not as forgetting.
This is the part that feels slowest and pays back most. An early product has very little traffic. A test on a page with a handful of visitors a week can take weeks to say anything. Stacking changes to go faster only means you never find out.
Page changes ship as A/B tests against today's page, even at tiny volumes, and every visit is counted. The read is honest about what it can say: "too early" is a common verdict, and the agent says it plainly.
Why it never merges, and never says done early
Two rules sit under everything.
The agent never merges a pull request. It opens the branch, writes the description, links the preview, and stops. The founder owns their main branch. An agent that merges its own growth experiments is an agent you will eventually turn off.
And it never calls anything done until a check from outside says it is live. Until then the work is "claimed". A merged pull request is not a shipped change if the deploy failed, the cache still serves the old page, or the new tag never made it into the HTML. So after a merge, the agent's server fetches the live page and checks for the change itself. Only then does the brief say verified.
The same goes for results. The agent reports what each shipped change brought once it can be read, in plain terms: "Worked", "Nothing yet". No green arrows on a chart nobody asked about.
What we got wrong along the way
A few things we learned by doing them badly.
Ranking without a diagnosis gave good moves for the wrong week. An SEO page is a fine move, but not the week your signup form is losing most of the people who reach it. Writing the diagnosis first fixed more than any change to the ranking.
Pre-written posts for communities got flagged. When we drafted finished comments for the founder to paste, moderators and readers spotted them. Now the agent drafts in points, the founder writes the sentences, and at least one line has to be theirs.
And a pull request a day is a ceiling, not a quota. Some days the right move is to wait for a test to finish. The agent says so instead of inventing work.
If you build with a coding agent and want to see the daily pick on your own site, that is what Shipfound does.
Frequently asked questions
How long should a pull request be?
Short enough to review in one sitting and focused on one change. A growth pull request is usually one page, one test or one fix. If it changes two things a visitor sees, split it, so each can be read on its own.
Why are pull requests needed?
A pull request puts a proposed change in front of the person who owns the code before it goes live. For growth work, that means the founder sees exactly what changes on their site, and nothing reaches visitors without their merge.
What makes a good pull request description?
It says what changes for the person using the product, why now, and how it will be read. For a growth change, that is the page, the number it should move, and the date the agent will check it. Link the preview so the reviewer can see it live.
What is the purpose of a draft pull request?
A draft pull request shares work that is not ready to merge yet, so others can look early without anyone merging it by mistake. It suits a change waiting on a decision from the founder, such as copy they want to write in their own words.