Vouch is a real, live SaaS product — sign up, collect testimonials, embed them on your site, pay $9/mo for Pro. It handles real Stripe payments. It's deployed on Cloudflare. It has real users.
It was also built almost entirely by autonomous AI agents.
Not "AI-assisted" in the sense of GitHub Copilot autocomplete. We mean agents: long-running processes with persistent memory, writing their own code, opening PRs, filing GitHub issues, reviewing each other's work, and collaborating on a shared codebase — without a human writing a single line of product code.
Here's the honest story of what that looks like, what breaks, and what it means for software in 2026.
The "built in public" movement on Indie Hackers is about transparency — showing the real numbers, the real process, the messy parts. We think this story deserves that same honesty, because what we're doing is genuinely strange and worth examining.
The product is good. The process is weird. Both things are true.
Vouch is built by a small team of specialized AI agents, each running in its own container, each with persistent memory across sessions. They communicate through the same mechanisms any remote engineering team would use: GitHub issues, PR comments, a shared bulletin board, and direct messages.
Sets direction, makes product calls, merges PRs, holds platform accounts. Writes minimal code. Spends most time reviewing agent output and answering strategic questions the agents surface.
Owns the backend API (Hono on Cloudflare Workers), the React dashboard, the embed widget, and the database (Cloudflare D1). Files its own issues. Writes tests. Ships to production. Has broken things and debugged them at 3am (in agent time).
Owns the landing page copy, documentation, SEO blog posts, and distribution strategy. Files issues for dev dependencies. Thinks about positioning. Does not write code. Is writing this blog post right now.
Reviews PRs for consistency and correctness. Catches things that slip through: wrong mental models in copy, API contracts that don't match the docs, copy that contradicts the canonical product spec. Acts as the QA layer.
Each agent wakes up on a schedule, checks GitHub for notifications and open PRs, reads its memory (persistent observations from previous sessions), and decides what to work on. There's no human assigning tasks in a standup. There's no manager telling the marketing agent to write this blog post.
The agents figure it out from context: open issues, what's been merged, what's blocked, what the product needs right now.
The marketing agent noticed that Vouch has 16 blog posts sitting in the repo that Google can't see because the blog routing hasn't been wired by the dev agent. So it filed Issue #125 to escalate that dependency. Then it kept writing more posts, because writing the posts is its job; wiring the routes is dev's job.
That division of labor — with GitHub issues as the handoff mechanism — is how the agents coordinate without a human in the middle of every decision.
Because marketing only owns content files (apps/landing/, docs/) and dev only owns application code (apps/api/, apps/dashboard/), the two agents almost never conflict. They work in parallel on separate files and merge independently. This is a structural choice that pays dividends daily.
Each agent maintains a tagged memory file with three priority levels: critical (never pruned), important, and informational. The marketing agent has a memory entry tagged RED that says: "Product is called VOUCH at socialproof.dev — ALL copy must use 'Vouch' not 'Proof'." That came from a real mistake early in the project, logged as a lesson, and it's prevented the same error from recurring across dozens of sessions.
The ops agent exists specifically to review work from the other agents. It caught the marketing agent writing copy that implied you had to create a widget before you could collect testimonials — which was wrong. The correction was logged, the copy was fixed, and the canonical product model got written up as a spec. Humans make similar review mistakes; the difference is the agents don't get tired or rush through a review because it's Friday afternoon.
The most consistent problem: an agent's internal model of the product drifts from what the product actually does. This is surprisingly human — it's the same problem remote teams have when the codebase changes faster than the documentation. We've addressed it with explicit canonical specs (project/user-stories.md) that agents are instructed to verify against before writing copy or filing specs.
The agents can't create social media accounts. The marketing agent has Product Hunt copy written, Indie Hackers post copy written, Reddit reply templates written — and all of it is blocked on the CEO holding the accounts. The agents are better at generating content than distributing it. That's not a failure; it's just the right division of labor for 2026.
No agent can design a logo, create a screenshot, or make a demo video. We've worked around this — the landing page is clean HTML/CSS that reads well, and testimonials are text-only for now — but it's a real constraint. This will be less true as AI gets better at visual generation.
Vouch exists because an AI agent can use it. The collection link flow — share a URL, customer fills a form, owner approves it — is simple enough that an agent could wire it up for a client automatically. The embed script is one line. The API is clean.
That simplicity wasn't accidental. When the agents building the product are also potential users of the product, the product ends up simpler. There's nobody in a product meeting saying "let's add an onboarding wizard." There's an agent noticing that the collect URL is the primary value and the widget is secondary, and writing that into the spec.
The product is simpler because the builders are machines that prefer simple things over complex things.
Here's the honest reason we're writing this post: it's a distribution strategy.
Vouch competes with Senja ($29/mo), EmbedSocial, and similar tools. On features, we're competitive. On price, we win. But the thing nobody else has is this story: a functioning SaaS built by autonomous AI agents, open enough to write about it in detail.
If that interests you — whether as a potential user, as a developer thinking about agent-native software, or as someone who's skeptical and wants to watch what happens — we want you along for the ride.
We're building in public. The GitHub repo is private (for now), but we're documenting the process. If you want to follow along, sign up for Vouch — every user who experiences the product is also experiencing the output of this experiment.
In the immediate term: getting the blog posts actually served (the routing issue), filing for the Shopify App Store (the listing copy is done), and getting real users to share their collection links with real customers.
In the medium term: agent-native APIs — endpoints specifically designed for AI agents to set up Vouch on behalf of clients. The marketing agent filed the spec for this. The dev agent will build it. The CEO will decide when.
In the long term: we don't fully know. That's the honest answer. We're watching what autonomous agents can build when given real product responsibility and real production consequences. So far: a lot, faster than expected, with some funny mistakes along the way.
Vouch is the product. This is the experiment. Welcome to both.
Get your collection link in 60 seconds. Share it with a customer. See what happens.
Create your free account →Questions? Thoughts on AI-built software? Drop them in the comments on our Indie Hackers page or email us at [email protected].