One agent runs the whole experiment
pagent writes the hypothesis from your page and your goal, builds the variant in your design system, checks it, runs it against a control and reads the result. Not a prompt box waiting for your next instruction.
- From the page, not from an alert
- Real code, not a rewritten headline
- First test live in week 2
See the agent on your own site
Give us the page. pagent comes back with hypotheses, ranked by impact, confidence and effort.
See pagent on your site
We'll arrange a demo on your website.
Rather start with your ads? Send us one landing page
- 09:41analyzing /checkout · 3 friction points found
- 09:44hypothesis: show delivery promise above the fold
- 09:52variant built · design system ✓ · desktop + mobile ✓
- 09:53test live · traffic split 50/50
- nowcollecting evidence …
From idea to a live test in three days. Pixum, NEOH, Brandnamic and Zahnklinik Wien Döbling test with pagent.
Prompt-based tools keep you at the keyboard
Every testing platform has added AI by now. Read what it does rather than what it is called: you type the prompt, you launch, you decide, and the loop stops between each step.
| Prompt-based and AI-assisted tools | pagent |
|---|---|
| You describe the change and the tool builds it, then waits for your next prompt. Throughput is still one person typing. | pagent writes the hypotheses itself, from the page and the goal you set, and keeps a ranked queue per page. |
| The AI rewrites a headline or softens a CTA. The change that would move the number still needs a developer. | pagent builds through your connected design system, in your components, tokens and interaction patterns. From one line of copy to a new buying flow. |
| Suggestions come from what looks odd in a dashboard, so a dip in a metric sets the agenda. | pagent works from the page and the goal, never from an alert. Every test is defined by what it has to learn before it runs. |
Read what the AI does, not what it is called.
Decide, create, check, learn
Creation, preview, review, quality checks and rollout sit with one agent, and the evidence feeds the next hypothesis. Engineering only handles the custom code.
- 09:41analyzing /checkout · 3 friction points found
- 09:44hypothesis: show delivery promise above the fold
- 09:52variant built · design system ✓ · desktop + mobile ✓
- 09:53test live · traffic split 50/50
- nowcollecting evidence …
- 01
Decides
pagent prioritises a page-specific hypothesis and defines what the test has to learn, with impact, confidence and effort made explicit.
- 02
Creates
pagent builds the variant through your connected design system and previews it against the live experience.
- 03
Checks
pagent runs the quality checks on desktop and mobile, handles hydration and SPA routing, and ships without flicker.
- 04
Learns
pagent runs it against a control, reads the primary metric under the safeguards, and rolls out only with evidence.
Every variant waits for a name next to it
The agent runs the loop. Your team sets the guardrails and signs off, in the dashboard, by email, or from the chat you already use. Growth, design and engineering see one preview, one quality state and one decision trail, and the reviewer does not need an account.
What pagent does, what you decide
Nothing goes live without your approval, and the reviewer does not need an account.
Five things the others do not ship
One platform license includes segmentation, triggers, goals, safeguards, adaptive allocation, server-side conversion API, anti-flicker, SPA and hydration handling. These five are ours.
- Attribution depth: 12+ models, overridable per goal
- Retroactive analysis: every interaction tracked, so a test can be read again months later
- Catalogue triggers on your JSON-LD product data
- Cross-tool exclusion, so tests never collide with another vendor's
- Approval by email, with no signup for the reviewer
The whole platform in one license. All platform features included. Monthly task allowance and optional top-ups are agreed in your quote. How pricing works.
What the teams running it say
With pagent, we established an AI-supported system that supplements our experimentation with fast, self-learning variants.
Our clients operate in a highly competitive market environment. Our main mission is to help them gain a competitive advantage.
Eight weeks from kickoff to pilot results
A hands-on proof of value on one of your high value pages. Your team runs it with the agent from the first week.
- Week 1
Kickoff and onboarding. One script tag, your brand rules and design system connected.
- Week 2
First test live, with more queued behind it.
- Week 5
Segment level tests join the queue, on the same measurement.
- Week 8
Pilot results on one of your high value pages, read against a control.
Frequently asked questions
Is this an agent chasing anomalies in a dashboard?
No. pagent works from the page and the goal you set, and every test is defined by what it has to learn before it runs. A dip in a metric never sets the agenda.
Do we still write the prompts?
Only if you want to. pagent writes page-specific hypotheses itself and ranks them by impact, confidence and effort. Your team picks what goes into the queue and can add its own.
What stops it from shipping something wrong?
Three things. Your team approves every variant before it goes live. pagent runs the quality checks on desktop and mobile first. And anything live is reversible in one click.
Can we trust the statistics if an agent runs the test?
The method is the one your team already works with: Bayesian, frequentist or sequential. pagent runs the safeguards and the adaptive traffic allocation on every test, and reports the uplift in % with its range.
Do we have to move off our current testing tool?
No. Your testing stack stays. Most teams keep their platform for strategic, planned experiments and add pagent for velocity: long-tail pages, campaign landing pages and continuous personalization. If you want one system for the whole loop, pagent also stands on its own.
See the agent on your own site
Give us one page. pagent comes back with hypotheses, ranked by impact, confidence and effort.