The auto-loop.
Never stops.
Watch → Research → Rank → Ship → Measure → Repeat. Eyepup's six-step autonomous-analytics cycle. An AI agent profiles every visitor, surfaces the highest-impact fix, hands it to your coding agent, then re-evaluates the next visitor after deploy. Inspired by Karpathy's autoresearcher loop — applied to your product's growth.
Watch
Every visitor, every event, every session — captured via PostHog.
Recorder agent ingests session events every 60s. Behavioural signals (rage clicks, dead clicks, scroll depth, errors, journey shape) hit Postgres before the visitor leaves.
Research
An LLM reads the session like a private investigator.
Profiler agent ships an rrweb-rendered MP4 + summary stats to Gemini 2.5 Flash via OpenRouter. Output: heat score (0-100), persona, intent, blocked-by reason, recommended action.
Rank
Friction patterns sorted by user impact, not feature count.
Pattern-finder agent clusters profiles into 3-25 stories per team. Pattern-compactor merges semantic dupes. Top 3 surface on /improve.
Ship
Your coding agent pulls the prompt and writes the diff.
Each pattern row carries a paste-ready prompt with the WHY, the file path guess, and the visitor evidence IDs. Claude Code / Cursor / Codex implement, you review.
Measure
Re-evaluate the next 100 visitors. Did the pattern shrink?
Pattern row gets marked 'fixing' when you click ship. After deploy, Eyepup re-profiles new sessions on the changed page and shows users_count delta on the next tick.
Repeat
The loop never stops. The agent gets sharper every cycle.
Resolved patterns feed back into pattern-finder's prompt as PRIOR PATTERNS so it stops re-suggesting fixes you already shipped — and flags regressions if friction returns.
The loop is the product. Sign up, point Eyepup at your site, and your first paste-ready fix lands within the hour.