# Vercel Expands 'Agentic Infrastructure' with Stakpak Acquisition > Vercel acquires Stakpak, boosting its agentic infrastructure strategy with autonomous DevOps capabilities. > By Dave · 2026-07-24 > Source: https://otf-kit.dev/blog/vercel-agentic-infrastructure Vercel just bought Stakpak — the Cairo-born team behind one of the first open-source autonomous DevOps agents — and the framing deserves more credit than the usual "AI buys AI" coverage. George Fahmy's line in the announcement nails what's actually shifting: "Vercel is no longer just the frontend cloud." The company is repositioning itself as the substrate for agents that build, deploy, and operate other agents. That sentence is doing real work, and the deal's structure confirms it. For two years Stakpak bootstrapped out of Cairo, then closed a $500,000 round in early 2025 — P1 Ventures led, with Digital Currency Group, 500 Global, and Instabug co-founders Moataz Soliman and Omar Gabr participating — while targeting $100k ARR with an engineering team kept in Egypt. The fact that Vercel moved on a team this small, this fast, two days after acquiring Ethiopian-founded Better Auth from developer Bereket Engida, says more about Vercel's roadmap than any roadmap doc could. The thesis: **agentic infrastructure is the bet, the deal is an acqui-hire for the founding team and the open-source projects, and the interesting question isn't whether the bet is real — it obviously is — but what survives the model churn underneath it.** ## What "agentic infrastructure" actually means Strip the buzzword and it describes a specific operational shape: agents deploy, developers build and run their own agents, and infrastructure is managed by agents themselves. That's the reframe Fahmy used announcing the deal. Not "we host your app" — "we host the agents that build, deploy, and operate your app." That distinction matters because the previous generation of "AI in DevOps" was assistive: copilots suggest commands, dashboards flag anomalies, CI pipelines get a smarter step. Agentic infrastructure is substitutive. The agent isn't advising the human — it's running the loop. Stakpak was one of the first teams to ship that loop as an open-source product, and Vercel now owns the team that wrote it. ![a developer pokes the agent harness, the harness shells into a sandboxed cloud account, ev](https://cdn.otf-kit.dev/blog/vercel-agentic-infrastructure/inline-1.png) ## What Stakpak shipped that Vercel is buying Three concrete artifacts, all open-source, all of which remain available after the deal: 1. An **open-source agent harness** — the runtime shell around a DevOps-capable model. Tool registration, credential handling, sandbox boundaries, the boring substrate that turns "LLM that knows `kubectl`" into "agent that can actually deploy your service without leaking prod secrets." 2. A **production-ready DevOps agent** — built on the harness, designed for the long-tail of deploy / roll back / scale / fix tasks that eat a senior engineer's week. 3. **AI-enhanced security systems** — the part that decides what the agent is allowed to do, and stops it when it shouldn't. If you're letting an agent touch production, this is the bit that has to be correct on day one, not bolted on after a postmortem. That third piece is the reason the deal matters. Most agent demos skip the security layer; Stakpak built it first and made it part of the same open-source bundle. ## How to actually use this today The acquisition closed, the open-source projects remain available, and Stakpak's team continues developing under Vercel. That means the practical move for a developer right now is to install the harness, point it at a non-production project, and watch what it does before you trust it with anything you care about. A reasonable first session: ```bash # Run the harness against a staging project you've already deployed by hand stakpak run \ --target=staging \ --task="scale ./services/hello to 3 replicas, run smoke tests, report" ``` ```bash # Inspect the run log later — this is the audit story stakpak replay --run-id=last ``` What to look for in that first run: - Does the harness **show you every command before it executes**, or does it run-then-tell? A good one prompts for approval on anything state-changing. - Does it **scope credentials per-task**, or hold one long-lived service account? - Does it **write a run log you can replay** when the agent does something weird at 2am? An agent you can't audit is an agent you can't keep. The point isn't to replace your CI on day one. It's to map where the harness stops being impressive and starts being a foot-gun. The teams that adopt agentic infrastructure successfully are the ones who map the failure surface in staging, not the ones who turn it on for prod on launch day. ## Where this fits Vercel's wider moves Two African-founded infra acquisitions within days is not a coincidence. Better Auth handles identity and authentication — the bit you need before any agent gets to do anything privileged. Stakpak handles the deploy / operate loop. Read together, the picture is: Vercel is buying the pieces needed to host an agent that touches your production stack with non-trivial privileges, and ship them as open-source-first tools that stay usable even if you don't host on Vercel. That's the part most coverage misses. Vercel could have kept Stakpak closed-source the way most acqui-hires end up. The deal terms — "open-source projects will remain available" — signal that Vercel wants the developer mindshare, not a walled garden. Whether that lasts is a separate question, but for now the artifacts are usable on any cloud. ## What doesn't change when the agent harness changes Here's the bit that survives the model churn underneath all of this. When Vercel ships a new agent harness, when the model behind it gets cheaper or sharper, when Stakpak's team open-sources the next iteration, the part of your product that your users actually touch stays the same problem it has always been: the same screen has to look right on web, iOS, and Android, and the component you ship on day one has to behave the same on day three hundred. That doesn't get solved by a better deploy loop. It gets solved by a single component API that produces the same output on every platform you ship to — so when the agent redeploys your app at 3am, the buttons still line up. That's the durable layer underneath the tool churn. Use the agentic infrastructure to deploy faster. Use a cross-platform component layer so the thing you're deploying doesn't drift across platforms the moment a model upgrade changes a default. The two are complementary, not competing. ![agent-managed infrastructure vs manually-managed CI — what survives both is the cross-plat](https://cdn.otf-kit.dev/blog/vercel-agentic-infrastructure/inline-2.png) ## What this enables Three concrete things the deal makes easier this quarter: - **Open-source agents you can audit.** The harness ships with a run log. You can replay what it did, diff it against your intent, and reject the run. Most proprietary DevOps tooling can't do that. - **A pricing pressure point.** Vercel competing with itself and with the open-source community on agentic tooling keeps the floor lower than it would be if the only option was a closed platform. - **A hiring signal.** If you wanted to work on agent infrastructure, the talent just concentrated under one roof. That changes who's reachable for the next round of open-source contributions. The first is the one to actually use today. The other two are worth watching. ## Closing Vercel bought Stakpak to own the team that wrote the first open-source agentic infrastructure, and Vercel kept the artifacts open on purpose. That's a real bet, executed cleanly, with concrete code you can run this week. The thing it doesn't solve — and never will — is the part of the stack your users feel. Use the agent. Ship the durable layer on top.