# Ideas That Outlive Their Tests

Hypotheses now live in their own workspace, with evidence, state, and outcome attached to the idea itself, not scattered across the tests that chased it. A new setup checklist and a few workflow upgrades round out this release.

Published: 2026-08-18

Category: feature

An idea used to live inside whatever test picked it up, and once that test ended, so did the record of why you believed it in the first place. Hypotheses are now a first-class thing in pagent, on their own workspace page. Every idea belongs to your site, can be scoped to a page and a trigger or to neither, and stays around long after any single test that chased it is gone.

A hypothesis carries its text, notes, where it came from, and whatever convinced you: screenshots, session recordings, survey exports, research documents. That evidence sits with the idea, so the agent building a test from it works from the same material you did. Its state (untested, testing, concluded, or parked and retired for the ones you've deliberately shelved) is derived from what's actually happening to it, never something you maintain by hand. Alongside the state, each hypothesis carries an outcome that pagent keeps in sync with its tests: pending, confirmed, rejected, contradicted, or inconclusive, with a short summary. When several tests target the same idea, the outcome accounts for all of them, and a contradicted outcome tells you two tests disagreed, which is worth knowing rather than averaging away.

Ideas connect to each other, too. From a hypothesis or from one of its tests, create a follow-up, a pivot, or a narrower version, and the workspace shows the whole lineage: what an idea grew out of and what grew out of it. Park an idea with a note when it's not the moment, retire it with a reason when the evidence says it's not a driver, and both read back exactly as you left them. Filter the list by state, outcome, page, and trigger, so "what's being tested right now", "what did we confirm", and "what did we shelve and why" are each one click apart.

Creating a test from a hypothesis is now the single path from an idea to a test, whether you start it from the workspace, from a review, or by asking in chat. The test carries the hypothesis with it, the agent gets the idea and its attachments as the briefing, and results flow back onto the hypothesis when it concludes. Rerunning an idea asks you to say why, so the record explains itself later. Chat and connected AI tools can work with hypotheses too: capture one from a conversation, attach files to it, park, unpark, or retire it, propose a follow-up, or turn one straight into a test.

## Know your setup is actually correct

A new Checklist page in settings runs about 40 automated checks across eight areas: the snippet itself, content security policy, bot protection, consent, goals, server-side conversions, single-page apps, and live verification. Each check shows the evidence it found and what to do about it, and you can re-check a single item, a section, or a whole page on demand. Anything you've handled deliberately can be confirmed or dismissed with a note, and the settings nav carries a badge with how many blocking items are left, so you always know how far you are from a clean setup.

## Tracking that reaches further

Test data can now flow into your team's own tools. A toggle under Integrations forwards test assignments and conversions into Google Tag Manager or Jentis, with no code required on your site: every enrolled test is pushed with readable test and variation names, conversions carry the goal, revenue, and properties, and visitors who opted out are never forwarded. The setup page also includes a copyable Shopify custom pixel, so orders placed on Shopify's checkout, which doesn't run your theme code, still get attributed to the visit that saw your test.

pagent also keeps watching a change after it applies, and reports when a page's own scripts render over it, with a new Overwritten Changes card on the errors page showing the share of sessions affected per change. Sites built with Astro are detected properly too, so changes wait for the page to finish hydrating before they apply. And preview links now travel with you: a preview stays active as you move around the site, including when a filter rewrites the URL, with a bar that names the test and variation, flags when the current page isn't targeted, and carries the way out.

A batch of smaller upgrades rounds out this release. You can rename any test from its options menu, and the name follows it into the tests list, the test page, the dashboard, chat, and your daily and monthly report emails, always alongside the test number. Chat has a searchable history panel for picking up a previous conversation, shows readable cards with an Open button for creating tests, capturing hypotheses, sending to review and lifecycle actions, and a tool waiting for your approval still works after you reload the page. Upload a Figma file on its own settings page and the agent reads your design variables, so generated changes use your colors, spacing, and type. The extension loads your site's test data once per session instead of per page, so moving around the site no longer means a wait or a flicker, with a Refresh data option when you want it, and it now evaluates JavaScript trigger conditions the same way the live snippet does. Triggers can gate a test on a specific page, not only site-wide. Every test shows its statistical setup in a dialog, the variations table explains each column on hover, and you can open a test in the editor straight from its review, where the entry button now reads Review Test. Conversion goal suggestions lead with the plain-language reason instead of the technical selector, SDK error reasons on the errors page are written in plain language and can be drilled into, personal access token settings gained a Select all for organisations and permissions, and your organisation view is remembered between visits.

[HTML](https://www.pagent.ai/changelog/ideas-that-outlive-their-tests)