Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

The Code Is Done. The Business Is Just Starting: Replit Agent Can Now Buy Supplies and Run Your Ads

On 12 November 2026, Replit gave Replit Agent two new jobs: buying supplies for your business and running your Google Ads. Purchases go through Stripe's Link wallet and Visa's rails; ads go through Google Ads. The common thread: the agent can only prepare the work — money moves only when the founder approves, with checks enforced in code, not left to the model.

A young woman seals a cardboard shipping box with a red tape gun in a bright home office, shelves of shipping boxes behind her, a laptop on the desk

In short: the code is done. The business is just starting

On 12 November 2026, Replit gave Replit Agent two new jobs on its official blog: buying things for your business, and advertising it. The first is called Agentic Commerce, paid for through Stripe's Link wallet and Visa's payment rails; the second is Agentic Ads with Google, which lets you set up a Google ad campaign without leaving a Replit chat. Neither job involves writing code, and both matter directly to the people who write code: for an independent founder who used an agent to build a product, the day the product ships is the day the spending really starts.

This is worth writing about not because Replit added another feature, but because it pushes the boundary of what a "coding agent" is. For the past two years, coding agents have competed on who writes code faster, fixes bugs more accurately, and sustains longer tasks. Replit just changed the contest: instead of competing on writing code, it is competing on who can take over the chain of work that turns a product into a business. Buying supplies and running ads happen to be the most tedious, most expensive, and most error-prone jobs in a small business. Whether an agent can do those two jobs well says more about whether it is a teammate than whether it can write one more feature.

The first new job: buying supplies — the agent shops, fills the cart, you approve

Start with the buying job, because its workflow states Replit's idea of "control" most clearly. You describe what you need in a Replit chat. The blog's example: "find 50 shipping boxes that fit this product, under $75 including delivery." The agent uses what it already knows about your business to find options, then a browser subagent navigates the merchant's website, picks the product options, fills the cart, prepares checkout, and returns the purchase details to the agent for your approval. You confirm the final total in Link before anything is charged.

Four details in that flow deserve to be read one by one. First, the agent shops on the merchant's own website, using a browser provided by Browserbase, so Replit needs no separate integration with each store — that decides how widely this can spread: in principle, any website where you can place an order is somewhere the agent can go. Second, the whole purchase is coordinated by the agent, but the last step always stops at checkout to wait for you: the blog says the agent "stops at the checkout for approval." Third, payment splits across Link and Visa: Link is Stripe's consumer network, with more than 300 million users globally — you pick a saved payment method inside it and approve the final total, and your payment details are never handed to the agent; Visa supplies the payment infrastructure and capabilities designed for agent-initiated transactions. Fourth, and the point Replit repeats in the blog: the agent "never gets open-ended permission to spend."

The blog also spells out the safety measures during a purchase, and they read like a reply to skeptics. The browser session is isolated, and shell access and general browsing are disabled during payment; approval checks are "enforced in code, not left to the model." In other words, "should this money be spent" is not the model's call at all — it is the program's. The model can only prepare; it cannot approve. That distinction matters, because it answers the first question everyone asks when they hear "an agent will spend money": what if the model gets it wrong? Replit's answer is that the decision never passes through the model.

But the launch scope starts narrow. Agentic Commerce is available for Core and Pro users, starting in the US, and limited to one-time purchases. Regular restocking and recurring subscriptions — the most common purchases in a real business — are outside the first release. Replit's documentation leaves a door open: you can also use Replit to purchase fixed-price subscriptions through Link, where the approval covers recurring payments rather than a single charge; but a supported subscription must have identical fixed initial and renewal totals, while trials, introductory prices, usage-based billing, prorations, and variable renewal totals are not supported, and a purchase cannot proceed if its recurring terms cannot be verified. The blog and the docs do not say exactly the same thing here — the blog says "limited to one-time purchases," the docs say fixed-price subscriptions work too — and a careful article should print that discrepancy as it is, not pick one version for Replit. What it can actually do in practice is a matter for the docs and the product itself.

The docs carry two more reminders worth pulling out on their own. First: "Connecting Link does not approve a purchase or charge you" — connecting Link is not approving a purchase, let alone being charged; payment approval is a separate step. Second: "Never paste payment card details, passwords, or verification codes into chat." Those two lines, written for users, also define the feature's boundary from the other side: Link is the door between you and your money. The agent can only push the cart up to the door. You are the one who opens it.

The second new job: running ads — the agent writes the copy and sets up the campaign, you approve the launch

The other new job is advertising. Replit Agent can now run Google ads for you inside Replit, across Search, YouTube, Display, Discover, and more, without leaving the platform. The flow is: connect or create a Google Ads account in chat, and the agent analyzes your website or app content, drafts ad copy, generates images, and recommends daily budgets; you review, approve, and launch — the blog says the whole thing takes "minutes."

The blog's advertising example is equally concrete: a founder selling bid-writing software might write, "Get us in front of contractors who bid on public projects. I'm willing to spend $40 a day for 30 days." On that wording, the 30-day budget is $1,200. You set the campaign goal, targeting, budget, and landing page, and the agent guides you with recommendations — even if you have never run ads before. The subtext is that the target user is not an advertising veteran, but a founder who has never run an ad in their life.

The most valuable line in the ads half of the blog hides in the middle: "Because Agent built your app, it can also add the conversion tracking that ads depend on." Anyone who has run ads knows the ad itself is not the hard part — knowing what the ad actually brought in is. Clicks alone are only the first half; conversion tracking watches sign-ups and purchases, the second half. Since the agent wrote your app's code, it knows where the sign-up button is and how the purchase flow works, so adding tracking is a small job for it and often the job a human keeps putting off. That detail is the most "agent as teammate" moment in the whole post.

The limits and thresholds are written more carefully in the docs than in the blog, and you only see the full picture by reading them. Agentic Ads with Google is available for Pro users worldwide from the day of the announcement. A new Google Ads account is created and managed by Replit, and ad spend is deducted from your Replit credit balance, with auto-reload able to charge your payment method to top it up; you can also connect an existing Google Ads account, whose billing arrangement stays separate. Only Performance Max campaigns are supported: Google combines your headlines, descriptions, and images into ads across its channels, and decides which combinations to show and where, based on your campaign goal. A Replit-created account can run one campaign at a time. The launch is for personal Workspaces, not Enterprise Workspaces, and Google promotional credits are not included in this release.

Three sentences about the budget in the docs have to be read verbatim. First: creating a campaign through Replit requires an average daily budget of at least $15 USD. Second: an average daily budget "is not a hard daily spending cap" — Google may spend more on some days than others. Third: "Ask Replit to confirm how and when your campaign will stop. Do not assume the campaign will stop automatically." Taken together, the meaning is plain: you set the budget, but the pace of spending is not the pace you imagine; when you want to stop, confirm it actively instead of assuming. The docs also warn that Google's reporting is delayed — usage and charges may appear after the activity occurred, including after you pause a campaign — and that Replit usage for preparing and managing ads is billed separately from the ad spend itself.

One more easily missed reminder sits in the docs' "Launch limitations" section: "A Mission budget is not an advertising spending cap." That line is a footnote to something else from the same week: Replit Missions, launched on 17 November, gives a swarm of agents a goal and a budget and lets it plan, build, and test on its own. Both features carry the word "budget," but they are not the same budget. Treating the Missions budget as the ads cap, or the ads budget as the Missions cap, will get your accounting wrong.

Why Replit: the agent's job list is spreading from "writing code" to "running a company"

To understand why these two jobs appear now, look at how many jobs Replit Agent has already taken on. In September 2024, Replit Agent launched, which Replit calls the first generally available coding agent: describe software in your own words and it gets built. That step opened the door to a new class of founders — people whose ideas had been held back by the high upfront cost and capability required to build software. Once the idea gets off the ground, a sprawl of new work follows: forming a company, keeping the app running, getting found, buying supplies, reaching customers. The blog's own words: "each one is a new barrier to growth," even for engineers who can build the product but don't know how to find customers.

Replit offers one number to show how far down that road it has already gone: since May, the number of entrepreneurs connecting their Replit-built businesses to Stripe for monetization has grown ninefold. Mind the denominator: the number counts "entrepreneurs connecting businesses to Stripe," not revenue, transaction volume, or retention. Ninefold sounds large, but the base is not published, so read the number for what it can and cannot say: it can say more people are building businesses on Replit and starting to take money; it cannot say those businesses survived or made a profit. The blog gives no numbers for the latter two, and this article will not invent them.

Before these two new jobs, the agent could already help a business accept payments, incorporate, find prospects and enrich leads, monitor app health and security, and optimize for SEO. Add buying supplies and running ads, and the agent now covers most of the chain from shipping a product to finding customers for it. The blog gives this direction a name at the end: the Self-Driving Company — agents take on more of the work so people can focus on strategy, ideas, and creativity. Replit states the ambition plainly: "help anyone turn an idea into a thriving business." That is a vision, not the current state, and the two should not be confused.

The direction also explains why Replit, rather than someone else, is doing this first. Its users already include large numbers of independent founders who built a product first and then figured out how to turn it into a business; the agent is the tool they built the product with, so it already knows what the product is, who it is for, and what it looks like. Letting the same agent buy supplies and run ads saves not one operation, but the work of "introducing your business to a new tool all over again." The blog's buying line — "using what it already knows about your product" — is exactly that logic: the agent does not need you to explain your product again before it shops, because it is the one that built it.

Two founder stories, and how to read them without being led

The blog ends with two founder stories, and they should be read with the premise "these are cases the vendor picked" kept in mind. The first is Jon Cheney: after a development shop quoted him $105,000 for an app he had been thinking about, he built the website, admin tools, and product on Replit in a weekend for under $400. The business made six figures in its first six weeks. The second is Austin Eckman: he paid a dev studio $50,000 for a first version he couldn't use. He now runs Crafting Table — teaching electronics and selling physical kits — from one Replit project, including an ads engine he built himself. In 30 days, he brought in $170,000, with a return on ad spend of roughly 4X.

Both sets of numbers come from Replit itself, with no independent verification, and the blog gives no further detail: whether the six-figure revenue is gross or net is unknown, how much ad spend went into the $170,000 is unknown, and how the roughly 4X return was calculated is unknown. What the two stories can show is only one thing: the path Replit wants you to believe in has really been walked by someone. They cannot show how easy the path is, how many people walked it, or what the failure rate is. This article includes them because they are part of the blog; but their evidence level is "two success stories picked by the vendor," not proof that independent founders generally make money on Replit.

The two quotes from Stripe and Visa should be read the same way, as statements from partners. Jay Shah, Business Lead for Link at Stripe, said: "A founder's to-do list doesn't end when the product is built. With Link agent wallet, Replit Agent can help buy the supplies the business needs, with every purchase approved by the founder." Rubail Birwadker, Global Head of Growth at Visa, said: "Replit is helping turn agentic commerce into something tangible for small businesses. As AI takes on more of the work, Visa provides the trust, control, and safeguards that help ensure agents transact within the boundaries people set." Both are partners speaking on launch day — endorsements, not independent evaluations. Their value is in telling you who built this together, not in proving how well it works.

Comparison: left, a timeline of the jobs Replit Agent has taken on — from writing code at its September 2024 launch, through payments, incorporation, lead finding, monitoring and SEO, to buying supplies and running ads on 12 November 2026; right, the approval flow for buying and advertising, where money only moves after the founder approves

Drawing the line with earlier coverage: this is not the first "agent spends money" article

Several earlier articles on this site touch "agents spending money," and they are easy to read as the same story, so the boundary is worth drawing. The first is Klarna's AI assistant buying books and running errands (around cycle218): that one is consumer errand-running, an assistant buying things and handling chores for you personally. The second is LangChain's payment rails and the x402 settlement layer (October): that one is protocol and infrastructure — how agents settle on-chain and pay in stablecoins — written for developers building the plumbing. This article is a third thing: a business tool for founders, where the agent buys supplies for your business and runs ads for it inside Replit, with money moving over Stripe Link and Visa's real payment rails and approvals hard-coded. All three share "agents starting to touch money"; they differ in whose money, where, and through which rails.

Two other Replit articles of its own also need boundaries. The 4 November piece on Replit's new sandbox engine is about the security foundation: each agent gets its own Linux kernel, with 96% of kernel vulnerabilities blocked — its keywords are isolation and safety; this piece is about two new business jobs on top of that foundation, and its keywords are spending and customer acquisition. The 17 November piece on Replit Missions is about giving a swarm of agents a goal and a budget and letting it plan, build, and test by itself — its keyword is "multi-agent collaboration for development"; this piece's keyword is "one agent spending money for the business." The Missions budget and the ads budget are not the same thing, and the docs' reminder about that is written precisely for people who would confuse them.

One more easily confused piece is October's Chroma joining Replit: that one is about an acquisition and the context layer, with Chroma Cloud shutting down within 60 days and ChromaDB staying open source — it has no direct connection to new agent features. This piece is not about acquisitions, only about features launched on 12 November. Read separately, Replit's last two months come into focus: Chroma joins at the end of October to strengthen the context layer, the sandbox gets a new kernel at the start of November to strengthen the security foundation, and mid-November brings Missions plus buying and advertising in quick succession — foundations first, features second, and the direction is from "helping you write code" toward "helping you run a company."

For people using agents to run a business: four things to settle before you start

Four practical points follow from this blog. First, test with a small order. The blog's example — 50 shipping boxes, $75 including delivery — is the right scale: the first time you let the agent spend money for you, pick an order you can afford to get wrong, walk the whole loop of "describe it, review the options, approve the total, check the bill," and confirm that Link approval, charging, and billing all line up before you trust it with a bigger purchase. The agent's ability to buy things should be proven on small money before it is trusted with large money.

Second, read the renewal terms before approving a subscription. The blog says one-time purchases; the docs say fixed-price subscriptions work too. With both wordings on the table, the safest move is to walk the docs' checklist before each approval: are the initial and renewal totals identical, is there a trial or introductory price, how often will the merchant bill you, is there an end date for the authorization, and how do you cancel? An authorization end date only stops permission for new payments — it does not undo a payment already authorized or cancel the merchant subscription. Cancelling has to follow the merchant's own process. Saying "I want to cancel" in a chat box does not count; the docs are explicit about that, and it is worth remembering.

Third, connect conversion tracking before you argue about how much to spend on ads. The agent can add conversion tracking for you, and that is where it is more convenient than other ad tools; but until tracking is connected, every dollar you spend only shows clicks, not sign-ups and purchases. The right order is: let the agent wire up conversion tracking first, confirm that sign-up and purchase events report correctly, and then set a daily budget — starting from the docs' floor of a $15 average daily budget, running a few days, and raising it on the data. An average daily budget is not a hard cap — Google may spend more on some days — so watch total spend, not any single day's spend.

Fourth, do not merge the two budgets into one. The Missions budget is a development-task budget; the ads budget is an advertising budget, and the docs' reminder exists because people will confuse them. When you want an ad campaign to stop, confirm how and when it will stop — do not assume it stops by itself. Google's reporting also lags, so charges can still appear after you pause. None of this is a product defect; it is the manual, already written down. People who read it will not trip over it; people who did not will mistake the rules for a malfunction.

Conclusion: the agent's next contest is not measured in lines of code

The line in this launch most worth noticing is not in the feature descriptions but at the top of the blog: what a founder has to do after the product exists — "forming a company, keeping the app running, getting found, buying supplies, reaching customers." Writing code is only the first item on that list, and for many people the hardest one — the agent solved that first. Now the agent is moving down the list: buying supplies and running ads. Neither job is glamorous, and no line of code in them is worth screenshotting, but they are the parts of a business where the money actually moves and where survival is actually decided.

That also changes the scoreboard for the coding-agent race. The old contest was benchmark scores, how many lines of code could be written in one go, and how complex a task could be finished independently; now there is another dimension: who can take over the string of jobs that comes after "the product is built," and do it in a way that earns trust. Replit's answer this time is: approvals hard-coded, money never passing through the model's hands, and the founder signing off at every step. Whether that answer really earns founders' trust will be decided by the bills and disputes of the first real users, not by launch-day blog copy. A blog can promise "owner in full control of the spend," but whether control is really in your hands is only settled when the first $75 box bill and the first $15 of ad spend reconcile.

Views 0Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates