# Vibe Agent
Vibe · 2026 · Lead Product Designer
An in-product agent that answers platform questions, turns a conversation into a report, generates campaigns, and recommends optimizations.

Source : https://albancarmet.com/project/vibe-agent

## Context

The advertising market was moving fast, and the platform had no agentic capability at all. The bet was to be among the first to ship one in ad tech. It was a product-led, top-down decision: no discovery phase, no user research. That was a constraint on every project at this company, where time and headcount never allowed for one. The build was sequenced like an MVP: first answering user questions about the platform, then reporting, then campaign creation and optimization recommendations, and finally contextual entry points spread through the product.

## Process

With no discovery to lean on, the only research we ran was a benchmark of tools solving adjacent problems, which we used as a source of patterns rather than answers. I led design across all four initiatives alongside a PM and a dedicated group of engineers. On campaign creation and on the contextual entry points, I used Claude Design as an exploration tool rather than a design partner: on the former, to see what graphical representations could be pulled out of the existing manual experience so we wouldn't reinvent the wheel; on the latter, to stress-test specific cases such as loading states, actions triggered from one module but landing on another, and how a user navigates between modules once the agent has answered. The decisions, and the arbitrations that came with them, stayed mine.

## Solution

Rather than bolting a chat window onto the product, every surface was designed to graft onto what already existed. Reporting turns a conversation into a presentable deliverable instead of another screen to learn. Campaign creation deliberately avoids rebuilding the manual form and its long list of parameters: the agent proposes a configuration, asks short questions when the request is too vague, and always ends on a recap that proves it understood before anything is created. Contextual 'Ask agent' buttons sit inside the manual flow, each carrying pre-filled prompts specific to its module.

## Results

Shipped inside campaign creation as a 50/50 A/B test over two weeks, the agent nearly tripled usage: the share of accounts using it went from 9.7% to 28.6% (+18.9 points, p < 0.001), and weekly usage held at about 20% versus 5% in the control group across both full weeks, so not a novelty spike. That result is what justified the rollout. An advertiser described the AI agent features as seamlessly integrated, making campaign management, analysis and optimisation effective and effortless. Four features were cut for build time, all decided with my PM and the engineering manager: predictive results (a pre-launch trend of key metrics based on advertiser and industry heuristics), pre-flight checks (automatic verification of payment method, tracking pixel and creative specs), a campaign budget rollup showing the split across strategies (replaced by a less visual module reused from manual campaign creation), and targeted editing by alias, where an 'Ask agent' button per module scoped the agent's edit to that module (users now describe the module in the main prompt). The alias idea has since partly come back in the manual flow, where targeting modules carry shortcut buttons to prompt that area; in the agent, what the missing aliases cost in round trips and tokens has not been measured yet. The team was an engineering manager, two engineers, a PM and me, over a month and a half from design to development.