Your customers already told you what to fix. Bellwether ships it, then checks it worked.

Bellwether reads your tickets, errors and analytics, finds the problem the most customers share, opens the pull request that fixes it, and measures the result.

Request early access See how it works

1,284signals this week

9problems found

$184kat risk in the leading one

Signals per week

Live demo, illustrative data. Click a sheep to inspect it, or click the grass to lead the flock. Pick a week to replay the season.

Where signals graze. The tools you already pay for.

Connect in minutes. Personal data is redacted before anything is stored.

TicketsZendesk, IntercomErrorsSentry, OpenTelemetryCallsGongAnalyticsPostHog, AmplitudeReviewsApp Store, G2CommunitySlack, Discord

One loop, from complaint to measured fix. Then around again.

Here is one real problem going all the way through: CSV exports failing for 12 accounts.

BellwetherYou

01
Ingest
Tickets, errors, calls and reviews flow in. Personal data is redacted first.
212 errors · 37 tickets · 4 calls
02
Cluster
Bellwether sees they are one problem and ties it to the accounts and revenue it touches.
OPP-118 · 12 accounts · $184k ARR
03
Spec
It writes the change, the acceptance criteria and the number that will judge it.
export_success_rate ≥ 95%
04
Code
Your coding agent builds it in a sandbox, behind a flag, and opens a pull request.
PR #218 · 148 tests pass
05
Ship
You review and merge. Rollout is gradual, and guardrails roll back anything that regresses.
1% → 100% · no guardrail tripped
06
Measure
The result is written back to the problem and the customers who asked are told.
62% → 96% · verdict: win
The verdict re-ranks the field, and the next bellwether leads.

The pens. Every problem has a place.

Problems move right as Bellwether and your team work on them. Open a card to see what is behind it.

Every pull request arrives with its reasons. No evidence, no PR.

Reviewers see who a change is for and how it will be judged before they read a line of code.

Every number is computed from stored evidence, not written by a model, so “12 accounts affected” is a query you can rerun.

Stream CSV exports instead of loading every row #218

MergedMaya merged 1 commit into main from fix/stream-csv-export

bellwether opened this for OPP-118

12 accounts ($184k ARR) cannot export tables over 10k rows. Exports load every row into memory and time out.

sentryTypeError in /export, 212 events in 7 days
zendesk“Export is greyed out for our big tables”
gong“We’d pay for a bigger plan if exports worked”
posthog31.4% of export_clicked never reach export_done

Judged by: export_success_rate, 62% → at least 95%
Guardrails: p95 API latency, error rate, auto-rollback

src/export/csv.ts+3−2

All checks passed · 148 tests, typecheck, guardrails armed
Maya approved these changes
Maya merged behind the flag export_streaming
Rolled out 1% → 10% → 50% → 100%, no guardrail tripped

bellwether measured this change after 7 days

Win+34

export_success_rate, 62% → 96% after merge · dotted line is the 10% holdout

Autonomy is a dial. You choose how far it roams.

Set it per type of change. Copy fixes can ship on their own while new features still wait for a person.

Bellwether does itBellwether drafts, you approveYou do it

Authentication, billing, data migrations and anything touching personal data need a person at every level, by default.

Three bets. Why Bellwether is different.

PostHog, Amplitude and LogRocket each shipped an agent that turns signals into PRs in 2026. Each one runs mostly on its own data.

Any source

Works on top of the tools you already have, and combines what customers say with what they do.

ZendeskSentryGongPostHog+7 more

Your coding agent

Writing code is a commodity. Bellwether decides what to build and checks that it worked, and your agent writes it.

Claude CodeCodexCursorCopilot

Every change judged

No evidence, no problem. No success metric, no pull request. Every shipped change ends with a verdict.

WinFlatRolled back

Shepherd’s log. Every loop ends with a verdict.

Wins, flat results and rollbacks are all written down, and the ranking learns from each one. Illustrative data.

Week 40
Billing page timeoutPR #204 · cache plan lookups
p95 6.1s → 0.9sWin
Week 39
Onboarding checklist copyPR #199 · shorter step titles
activation 41% → 42%Flat
Week 38
Bulk invite modalPR #192 · new layout
errors +3.1%Rolled back
Week 37
Password reset email delayPR #188 · move to a queue
tickets −64%Win

Join the first flock. Make the next thing you ship the thing they asked for.

We are working with a small group of design partners who have real users, real logs, and more feedback than time.

Become a design partnerhello@bellwetherlabs.dev

40	export async function exportCSV(query, stream) {
41	const rows = await db.select(query)
42	return toCSV(rows)
41	for await (const page of db.pages(query, 1000)) {
42	stream.write(toCSV(page))
43	}
44	}