The AI Spend Question Cross-Border Sellers Keep Avoiding
Every operator I talk to has quietly become an AI company whether they admit it or not. The listing copy generator, the review-summary tool, the ad-copy A/B tester, the customer-service bot bolted onto Zendesk — each one is a line item on a card that nobody reconciles against a P&L. So when a thread titled “How much do you spend on AI?” surfaces on Product Hunt, it lands differently for a Shopify brand owner than for a seed-stage dev tool team. The scrape I’m working from here is thin — a forum thread, a launch post, a reply — but it’s a useful mirror. It shows how the tooling layer underneath e-commerce is being rebuilt for agents, and why the rebuild matters to people whose actual job is moving units across borders.
What Murmell Is Actually Selling
The launch post is from Moss’Ab Mirande-Ney, co-founder of Murmell, a two-person team. The product is AVP — Agent Versioning Protocol — and it’s open source. The pitch is blunt: Git was built for humans, AVP is built for AI agents.
The origin story is the part worth stealing. Mirande-Ney cites Anthropic’s own disclosure that it moved its monorepo off Git after agents pushed it from roughly 2,500 merged changes a day to more than 30,000, using an internal system called Scion. The top Hacker News complaint, per the post, was that nobody outside Anthropic could use it. Murmell hit the same wall at smaller scale: two people, many coding agents from different providers, one monorepo. The bottleneck stopped being the agents and became the pull request. Every change waited on CI, rebased onto a main that had moved, and two agents editing the same file ended in a conflict someone had to settle by hand.
AVP’s answer: an agent opens a task, writes, and lands it. Each landing is proven against the exact head of main in seconds; if main moved, the task replays and is proven again. Every landing is still an ordinary Git commit, so GitHub and existing deploys don’t change. On their repo, CI runs for pull requests dropped from 104 a day to 4. In their benchmark with six to eight Claude Code agents on one repo, main went red under Git and Jujutsu and tasks silently went missing; with AVP, main stayed green and nothing was lost, in every run so far. The install is a single binary: avp up in any Git checkout connects Claude Code and Codex.
Why this is a developer-tooling story, not an e-commerce one — yet
Let me be honest about the gap. Nothing in this launch touches product research, paid acquisition, fulfillment, or returns. A Shopify operator running Klaviyo flows and a 3PL integration will not install AVP tomorrow. But the structural problem AVP names — many agents, one shared state, no reliable way to prove a change landed cleanly — is the exact problem every mid-size seller is about to hit as they wire agents into ops. The tool is early; the pattern is not.
The Problem Underneath: Agents Don’t Have a Merge Strategy
Cross-border sellers have spent two years buying agents. An agent that rewrites Amazon bullet points across 400 SKUs. An agent that reconciles supplier invoices against Amazon Seller Central reports. An agent that drafts Etsy listing variants. Each one is a separate subscription, a separate prompt, a separate place where state can drift. Nobody has a versioning story for any of it. When the listing-rewrite agent and the translation agent both touch the same product record, you get the e-commerce equivalent of a merge conflict: two conflicting descriptions live in the catalog, and the customer sees whichever one wins the API race.
That is why the Murmell thread is more relevant than it looks. The core insight — every agent action should be a provable, replayable, atomic landing against the current state — is a design principle sellers can demand from their vendors even if they never touch AVP. If your listing tool can’t tell you exactly what state it changed and prove the change against the current catalog, it’s a Git-era tool wearing an AI badge.
Where the math breaks
The benchmark numbers are the most quotable part of the launch and the least transferable. Six to eight Claude Code agents on one repo is a developer scenario. A seller running three agents across a 2,000-SKU catalog has a different failure mode: not merge conflicts on code, but silent overwrites on product data. AVP’s “proven on the exact head of main” guarantee doesn’t map cleanly to a Shopify product record that five apps can write to simultaneously. The principle transfers; the implementation doesn’t. Anyone who reads “main stayed green” and assumes their catalog is now safe is over-reading the claim.
How It Compares to the Incumbents
The launch explicitly benchmarks against Git and Jujutsu, and the framing is deliberate: both were built for human-paced collaboration, and both buckle when agent throughput rises. The Anthropic/Scion comparison is the strongest signal — a well-resourced lab concluded the same thing and built internal tooling rather than shipping it. Murmell’s bet is that the external market wants the open-source version.
For sellers, the useful comparison isn’t Git at all. It’s the agent-orchestration layer they’re already paying for. Tools like Zapier and Make wire agents to actions but offer no versioning or replay semantics — a failed run is a failed run, and you find out from a customer email. The AVP thesis implies a future where ops automation gets the same commit-and-prove discipline that engineering has had for two decades. That’s a real gap in the seller stack, and nobody in the e-commerce SaaS layer is filling it yet.
Why Amazon sellers should care more than Shopify ones
Shopify operators tend to run fewer, more curated SKUs with tighter human review, so agent chaos is survivable. Amazon FBA brand owners run wide catalogs, thin margins, and a marketplace that punishes listing errors with suppression and lost Buy Box. An agent that silently overwrites a title across 300 ASINs is a revenue event, not an inconvenience. The versioning discipline AVP argues for — prove every change against current state, replay if state moved — is worth more per dollar to an Amazon seller than to almost anyone else in e-commerce. I’d expect the first serious e-commerce adaptation of this pattern to come from the FBA tooling side, not the DTC side.
What Cross-Border Sellers Should Borrow From This
Three transferable ideas, in order of practicality.
First, treat every agent action as a commit, not a fire-and-forget call. Even without AVP, you can enforce this in your own stack: require agents to log the before-state, the intended change, and the after-state, and reconcile against the live catalog on a schedule. If a tool can’t do that, deprioritize it.
Second, measure agent throughput against human review capacity. Murmell’s 104-to-4 CI reduction is a throughput story, but the real lesson is that they instrumented the bottleneck. Sellers should instrument the same thing: how many agent-generated changes per day actually get reviewed, and how many slip through unreviewed. That ratio is your risk exposure.
Third, watch the open-source angle. AVP being open source and a single binary lowers the trial cost to near zero for anyone with a Git checkout. The reply from Haplo AI on the thread asks the sharpest question — how does AVP handle two agents touching the same lines nearly simultaneously, and does the later task replay cleanly or need human help? That’s the question sellers should be asking every vendor, not just Murmell. The answer to “what happens when two agents collide” is the whole ballgame.
The AI spend question nobody answered
The thread that kicked this off asked how much operators spend on AI. The scrape doesn’t give us the answers, only the framing that they were “gold.” That’s the honest state of the industry: everyone is buying agents, almost nobody is reconciling the spend against outcomes. AVP is a tool for the teams whose agent spend has already outgrown their ability to manage it. Most sellers aren’t there yet — but the trajectory is clear, and the sellers who build the versioning and reconciliation muscle early will be the ones who survive the throughput jump.
Where My Judgment Says It Falls Short
The launch overreaches in exactly one place, and it’s the headline: “Git was built for humans. AVP is built for AI agents.” Git is also built for durability, distributed trust, and a twenty-year ecosystem. AVP inherits Git’s commit format precisely because that ecosystem is too valuable to abandon — which is an admission that the replacement story is softer than the slogan. Mirande-Ney concedes that GitHub stays as a fallback storing the code and that they’re “slowly moving to our own client.” That’s a migration in progress, not a finished replacement.
Second, the evidence base is a two-person team’s own repo and a self-reported benchmark. No third-party validation, no disclosed pricing (the post says open source, but the commercial model isn’t stated), and the claim that “most teams will make the same move in the coming years” is a prediction, not a finding. For a seller deciding where to spend attention, that’s fine — it’s a signal to watch, not a tool to adopt.
Third, and most important for this audience: there is no e-commerce surface area here at all. No Shopify integration, no Amazon API story, no catalog-versioning feature. The transferable value is conceptual. Anyone pitching AVP to a seller as a direct fit is overselling.
What I’d Watch / Test Next
This week, do three things. First, audit your agent spend the way the Product Hunt thread implies — pull every AI subscription off your card statement and tag each one to a measurable outcome (listings changed, tickets deflected, ads tested). If a tool can’t be tagged, that’s your first cut.
Second, pick your highest-risk agent workflow — probably bulk listing edits or translation — and add a before/after reconciliation step, even if it’s a spreadsheet and a scheduled export. You’re building the muscle AVP argues for, without waiting for the tooling.
Third, follow the AVP thread and specifically the collision question raised in the replies. If Murmell answers it well and ships a client beyond the Git fallback, that’s the moment to take the pattern seriously for e-commerce. Until then, treat this as a well-argued design principle from a two-person team — worth reading, worth borrowing from, not yet worth betting your catalog on.





