# The Vibe-Coding Paradox ## Why cloud AI can't fix QA, and how local multi-agent VLMs change everything **Author:** Javier Alcivar — UX Strategy Lead Consultant at Globant **Version:** 1.0 · 2026 **Product:** Clikiti — AI cats testing your app --- ## 1. The flaw in first-generation AI QA The first wave of AI-powered QA tools inherited a pricing model from the era of hosted language models: every action an agent takes is a metered API call. That single design decision produces three structural problems that no amount of product polish can fix. **1.1 Costs scale against coverage.** Cloud AI-QA platforms commonly start around $500/month for roughly a thousand test executions, and bring-your-own-key tools move the same bill to your OpenAI or Anthropic account instead of removing it. Because each run has a marginal cost, the rational response is to test less — precisely the opposite of what a fast-shipping team needs. Coverage becomes a budget line rather than an engineering decision. **1.2 Test execution is stateless.** Most tools begin each run with no memory of prior runs. They do not know which flows broke last month, which forms are fragile, or which parts of the app changed. Every sprint restarts from zero, so coverage never compounds and the tool never becomes more valuable over time. **1.3 A single agent cannot find concurrency bugs.** This is the most important limitation and the least discussed. One agent, acting sequentially, cannot reproduce a class of defects that only exist when two or more users touch the same state simultaneously: inventory oversell, double-booking, coupon and credit reuse, race conditions and deadlocks, session and auth-token bleed. These are not infrastructure load problems — the server stays up. They are correctness failures in business logic under concurrency, and they are exactly the bugs that cost real money when they reach production. **1.4 The localhost gap.** Cloud platforms cannot natively reach `http://localhost:3000`. Tunneling through ngrok or Cloudflare Tunnels is possible but breaks the tight local feedback loop that makes fast iteration work in the first place. --- ## 2. The local agent paradigm Two things changed recently, and together they invalidate the assumption that serious AI QA must run in the cloud. **2.1 Vision-language models crossed the usability threshold.** Open-weight VLMs in the 7B–32B range now ground GUI elements accurately enough to drive a browser reliably. Feeding the accessibility tree rather than raw HTML keeps context small and avoids memory overflow, which makes quantized models on consumer hardware a practical substrate rather than a compromise. **2.2 The hardware is already installed.** A large share of developers already own a discrete GPU or an Apple Silicon machine with substantial unified memory — bought for gaming, video work, or local model experiments. That capacity sits idle most of the day. The consequence is a different cost structure, not merely a cheaper one. When inference runs on hardware the user already owns, the marginal cost of a test run is zero. Coverage stops being a budget decision. Test data never leaves the machine, which resolves the privacy and air-gap objections that block cloud QA in regulated environments. And the vendor has no cloud inference bill to pass through, so there is no incentive to meter usage. This is a moat rather than a feature. A competitor whose entire cost base is hosted-LLM inference cannot match zero marginal cost without dismantling its own business model. --- ## 3. Four personas, and why more than one matters Clikiti ships four agents with deliberately different behavioural strategies: - **Whiskers (curious)** explores methodically — hovers, reads alt text, checks links — and surfaces UX friction. - **Shadow (sneaky)** attacks edge cases: blank inputs, rapid clicks, back-button abuse, expired sessions. - **Paws (clumsy)** fat-fingers interactions, producing double-submits and duplicate writes. - **Tiger (aggressive)** pushes parallelism until business logic cracks. Behavioural diversity broadens the paths walked through an app, but the decisive argument for multiple agents is interaction. Running several cats at once creates cross-agent interference: two checkouts racing for the last unit of stock, two edits landing on the same membership row, two logins resolving against the same session store. That interference is the only reliable way to surface concurrency-class defects, and it is structurally unavailable to single-agent tools. A second benefit is speed. A pride of agents completes a full sweep in minutes rather than hours, which is what makes autonomous exploration viable as a CI gate instead of an overnight job. The framing matters too. Random "monkey testing" produces noise. Personas are curious and deliberate: they pursue goals, form expectations, and report the gap between the two. --- ## 4. The Brain: coverage that compounds Stateless testing wastes the most valuable artifact a QA system produces — knowledge about where a specific application tends to break. The Brain persists three things across runs: a session log of what each agent did and what happened, a synthesized map of fragile zones ranked by historical failure density, and a record of how the application's surface changed between runs. Later sessions read that history and bias exploration toward regions with a high prior probability of failure, while still reserving effort for unexplored areas. The practical effect is that the tenth run is materially better than the first, and a team's accumulated history becomes a switching cost that a stateless cloud tool cannot replicate by matching features. --- ## 5. Integration surface: MCP and Cat Cam **Model Context Protocol.** Exposing Clikiti as an MCP server lets the coding assistant already in the developer's editor request a test run and read back structured results. QA moves inside the authoring loop instead of sitting behind a separate dashboard. **Cat Cam.** Every finding ships with a short screen recording of the exact sequence that produced it, plus the agent's stated intent at each step. A reproducible clip converts an ambiguous bug report into an actionable one, which is where most of the human time in QA is actually spent. --- ## 6. Conclusion The first generation of AI QA was built on a metered cloud call, and inherited three consequences from it: costs that scale against coverage, no memory between runs, and blindness to concurrency-class defects. Local multi-agent inference removes all three at once — not by being cheaper, but by having a different cost structure and a different architecture. Zero marginal cost per run means test everything. Multiple concurrent agents mean finding the bugs that only exist when users collide. A persistent Brain means coverage that compounds instead of resetting. Not monkey testing — cat testing. --- *Clikiti · clikiti.com · MIT licensed*