AI product agents
Understand your users. Build what matters.
Install the SDK. Somar reads how people actually use your product — every event, recording, error and experiment — then writes the code, opens the PR, and tells you whether it worked.
Signals
The SDK captures what people actually do — events, recordings, errors, flag evaluations, survey responses, log lines. One weighted pool, and every tool that reads it is included.
Somar Code
Ask the agent what to build. It reads your funnel and edits the file that broke it in the same turn — then opens the PR, behind a flag, at 10%.
Rollouts
Every change ships with a flag and a metric chosen before it went out. Weeks later the list says whether it won, lost, or hasn’t enough data to say.
Live charts — signup to activation, updating as signals arrive from the SDK.
Masked recordings — watch the sessions that dropped. Inputs are masked as they are recorded, so the value never leaves the page.
All of it is the agent’s context — the same data you read is the data it reasons over before it touches a file.
Drop-off concentrates at step 2 — sessions that miss the skip control
rarely return. I’ve moved it inline and opened a PR, behind
onboarding_skip_inline at 10%, measured on activation rate.
- Shipped behind a flag at 10% — never un-flagged, so a mistake is a toggle
- Rolled to 50%, then 100%, watching activation rate the whole way
- Verdict: won — activation up, and the list says so without being asked
The metric is chosen before it ships, so the list can also say lost, not enough data, or stuck — a rollout nobody has touched in a month.
Understand your users. Build what matters.
Join the waitlist — we’ll email you when Somar launches.
Join the waitlist