Should PostHog Replay Vision automatically open pull requests for UI issues it detects, or only suggest fixes for humans to approve?

Introduction

Replay Vision, launched by PostHog in mid-2026, adds AI to Session Replay. Instead of humans watching hundreds of recordings, it scans sessions for signs of struggle, such as rage clicks, dead clicks, error loops and abandoned forms, summarises what went wrong, and can open draft pull requests for UI issues it detects. It is priced in credits, with 2,500 free credits a month and $0.01 per credit after that. The question is whether Replay Vision should open pull requests automatically for every issue it detects, or only suggest fixes that a human then chooses to turn into a PR. This is a real product trade-off: automation drives PostHog's "self-driving product" vision and makes the feature feel magical, but wrong or noisy PRs can damage trust with the engineers PostHog serves.

Step 1

Clarifying Questions (3 minutes)

• Context: What does "automatically open a PR" mean here?

• Why it matters: A draft PR in a sandbox branch is very different from a PR that requests review from the team.
• Hypothetical answer: Replay Vision opens a draft PR in the customer's GitHub repository, assigned to nobody, with evidence linked.
• Impact: The risk is noise and review burden, not code being merged without review.

• Business context: How does this feature make money?

• Why it matters: Automatic PRs consume more credits.
• Hypothetical answer: Each analysis and PR uses credits; automatic mode would raise credit usage.
• Impact: We must avoid optimising for credit revenue at the expense of value.

• User impact: Who receives these PRs?

• Why it matters: Small startups and large companies have different tolerance for bot PRs.
• Hypothetical answer: Mostly teams of 5–50 engineers.
• Impact: Repository noise affects the whole team, not just the person who enabled Vision.

• Technical: How accurate are current suggested fixes?

• Why it matters: Sets the baseline for trust.
• Hypothetical answer: About 45% of suggested fixes are accepted with small or no edits in early data.
• Impact: More than half of automatic PRs would need changes or be closed.