# Vibe Agent
Vibe.co · 2026 · Lead Product Designer · Project
Source: https://albancarmet.com/project/vibe-agent
An in-product AI agent that builds campaigns, reports on them in conversation and can be reached from every module.
In one line: Three surfaces, one continuous agent.
Team: EM, 2 engineers, PM and me · Duration: 1.5 months · AI: Exploration with Claude Design

## At a glance

- Problem: Ad platforms were starting to adopt agents and Vibe had none, while creating a campaign meant working through dozens of parameters.
- What I did: Designed three agent surfaces grafted onto the existing product: conversational reporting, campaign generation, and “Ask agent” buttons inside the modules.
- Result: In a two-week 50/50 A/B test on campaign creation, agent usage nearly tripled (×2.9) and held steady; the agent was rolled out to every account.
- Limit: No user research beforehand, and four planned modules were cut to fit six weeks of build.

## The company

Vibe is a self-serve advertising platform for connected TV. Brands of every size, from fast-growing direct-to-consumer labels to large advertisers, buy ads on the major streaming channels, target their audiences — including Walmart shopper audiences — and measure the sales they drive, the way they would on search or social.
- Sector: Ad tech · connected TV
- Customers: Thousands of brands, from DTC to large advertisers
- My time there: Since Dec. 2023 · Lead Product Designer

## The product: The platform

Everything on Vibe revolves around the campaign, and three steps come up throughout this page.
- Upload creatives: A creative is a video ad. Advertisers upload theirs to the platform.
- Launch a campaign: A campaign sets a goal and is split into strategies: who to reach, with what budget.
- Measure: The dashboard and reports show spend, reach and the sales driven.

## Process

AI came in where it helped explore, never where the call was made — and not at all on reporting.
- Benchmark: Agentic tools on adjacent problems: patterns to borrow, not answers. (without AI)
- Explore: Campaign creation and Ask agent: forms drawn from the manual flow, edge cases stress-tested — loading, actions landing in another module. (with AI, Claude Design)
- Design reporting: d3.js’s catalog of visualizations, adapted to the brand, with scaling rules from narrow column to full screen. (without AI)
- Ship & measure: Built by two engineers; a 50/50 A/B test on campaign creation, then rollout to every account. (without AI)

## The problem — Ad platforms were about to be reinvented by agents, and ours had none

Creating a campaign on Vibe meant working through dozens of parameters, reports took long to assemble, and platform questions had no answer inside the product.
The bet, made top-down and without a discovery phase: be among the first to bring agents to ad tech, as a differentiator rather than a feature.

## The approach — Graft the agent onto the product, don’t rebuild the product inside the agent

Users already knew the manual campaign flow. Rebuilding that form inside a chat would only have moved the same work into a narrower container.
So the agent proposes and the user reviews, on three surfaces each built on a task the product already handled.

## The trade-offs — Four modules cut to ship in six weeks

Every cut came down to build time, and the team agreed on the value of what was cut. What matters is what each one cost the user, and what makes up for it.

## What shipped — Three surfaces, one continuous agent

Each surface below replays the shipped flow as a playable prototype. I rebuilt these replays for this page with Claude Design, after the project: the product itself was built by the engineering team.

## Constraints

- A mature product, already learned: Users knew the manual campaign flow. The agent had to graft onto it, not duplicate it in a conversation.
- A generative flow can be confidently wrong: A campaign needs a complete, valid set of parameters, while requests range from very precise to very vague.
- Six weeks to ship: A month and a half from design to development: every module of the campaign panel had to justify its build time.

## Signals and research

No formal discovery: a product-led bet to be among the first to bring agentic features to ad tech. The decisions rested on these signals.
- Market — A window to take: Ad platforms were about to be reinvented by agents, and ours had none. Speed mattered more than completeness.
- Benchmark — Agentic tools on adjacent problems: A source of patterns, not of answers: it told us how others show a generated result, never what our users needed.
- Existing product — The manual campaign flow: Dozens of parameters and modules that already worked. Reusing them was both a constraint and a shortcut.

## Decisions and trade-offs

- Reuse the existing budget module instead of a dedicated rollup. Decided with: My PM and the engineering manager. Rejected: A “Campaign budget • rollup” — At the top of the campaign panel: the total, the implied daily pace, and color bands showing how the budget splits across strategies, with each strategy’s budget type (lifetime or daily). Kept: The manual flow’s budget module — A module reused from manual campaign creation. Less visual, but it let us ship the campaign summary much sooner. Why: Build time. The whole team agreed the rollup added value; it lost to a module that already existed. What it cost: Seeing at a glance how the budget splits across strategies, and which ones run on a lifetime or a daily budget. What makes up for it: The information is still there, in a format users already knew from the manual flow.
- Edit through the main prompt, without targeted aliases. Decided with: My PM and the engineering manager. Rejected: Targeted editing by alias — Each module of the panel had its own “Ask agent” button, icon-only by default and labeled on hover. It dropped the module’s name into the prompt as an alias, so the agent edited that module only, then showed what changed with an undo. Kept: One prompt for everything — Users edit the campaign from the agent’s main text input, describing in their own words the module they want to change. Why: Development time. Aliases were about comfort and precision, not about making editing possible. What it cost: Context handed to the agent for free. Without it, the user has to name the module, and a misread prompt means a round trip — plus, on our side, likely more tokens spent on corrections. What makes up for it: The model understands intent well enough that most requests land. How much the missing alias really costs is still unmeasured.
- Postpone predictive results. Decided with: My PM and the engineering manager. Rejected: Predictive results — An early concept: a trend of the campaign’s key metrics before launch, based on heuristics from the advertiser and its industry. Why: Development time, like the previous two. What it cost: An idea of what a campaign would deliver before committing the budget.
- Postpone pre-flight checks. Decided with: My PM and the engineering manager. Rejected: Pre-flight checks — A module that automatically checked the essentials for a campaign to run properly: a payment method on file, the tracking pixel installed, creatives meeting the specs. Why: Development time, like the previous two. What it cost: Catching a blocker before launch rather than after: a missing payment method or pixel still surfaces later.

## What shipped

- Conversational reporting (Turning a conversation into a presentable deliverable.): problem: Building a report is long and tedious work, and the agent could render results either in a narrow column or in full screen. Every visualization had to scale cleanly between the two without degrading into an unreadable chart. Solution: We designed the visualizations to hold their meaning at both widths, so the agent could answer in place or expand without the user losing the thread. A number of visualizations I had designed were cut when roadmap priorities shifted. The scaling rules were what made the shipped subset survive that cut.
- Campaign generation (Reviewing a configuration instead of filling a form.): problem: User requests range from very precise to very vague, while a campaign needs a complete and valid parameter set. Rebuilding the manual form inside the conversation would have removed the whole point of the agent. Solution: The agent proposes a configuration and presents it as something to review. When the request is too vague, short question forms appear inside the conversation to close the gaps, and a recap checkpoint proves the request was understood before the campaign is created.
- "Ask agent" entry points (Reaching the agent with context already attached.): problem: A "global" agent button already existed. Adding more access points risked being redundant noise rather than help, and raised concrete cases: what loads, what happens when an action fired from one module affects another, and how the user gets back to their work afterwards. Solution: Each button sits in the module it serves and opens with pre-filled prompts written for that module. The agent knows which part of the experience is in play before the user types anything, which makes both the visibility of agentic features and the quality of the answer a design outcome rather than a model one.

## Results

Shipped inside campaign creation as a 50/50 A/B test, the agent nearly tripled usage — and the usage held.
(A/B test · 2 weeks · campaign creation)
- ×2.9: Accounts using the agent
- 9.7% → 28.6%: Agent usage rate, control vs test
- +18.9 pts: Statistically significant (p < 0.001)
- ~20% vs ~5%: Weekly usage, test vs control, over both full weeks
Weekly usage stayed level across both full weeks: this is not a novelty spike. That is what justified the rollout.
Customer quote: « The UX is intuitive and robust — and the fast-developing AI Agent features are seamlessly integrated to make campaign management, analysis and optimisation both effective and effortless. » — A Vibe advertiser, shared by their CSM

## Still open

We never measured what the missing aliases cost. The idea has partly come back since, in another project: in the manual flow, targeting modules now carry shortcut buttons to prompt that specific area. In the agent itself, the cost of round trips caused by a misread prompt, and the tokens they burn, is still unmeasured.

## What AI did on this project

Claude Design as an exploration tool, not a design partner. Two of the three surfaces used it; reporting was designed without it.
- Campaign creation: To see which visual representations could be pulled out of the existing manual flow, so we wouldn’t reinvent the wheel.
- Contextual entry points: To stress-test edge cases: loading states, an action fired from one module landing in another, getting back to work once the agent has answered.
- Reporting: No AI: inspired by d3.js and its catalogue of ready-made visualizations, adapted to our branding, then scaling rules for each one between a narrow column and full screen.
- What stayed human: Every decision and every trade-off on this page.
