Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app

Nov 2, 2026•wuweiweiwu•View Original

TL;DR Highlight

Mo, an AI QA agent that explores your app, found about 269 labeled bugs on QA Bench versus 86 for the next best.

Who Should Read

Web or mobile developers shipping fast with coding agents who struggle to write and maintain E2E tests

Core Mechanics

Mo explores your app from a single prompt
- Mo is an AI QA engineer that explores the whole app from a prompt.
- You review the findings, recordings, and repro steps.
Mo found far more bugs on QA Bench
- It ran 60 minutes on each of 20 areas across Gitea, Metabase, Twenty, and Saleor.
- Mo found about 269 labeled bugs, and the next best agent found about 86.
- The comparison was against Codex, Claude Code, and Devin.
A separate agent reproduces each bug
- A separate agent reproduces each functional bug before Mo reports it.
Tests live in your repo as YAML
- Steps and assertions are written in plain-English YAML.
- The CLI exit code can serve as a required CI check.
Tests repair themselves when the UI changes
- Auto-heal re-finds controls during a run.
- Permanent healing proposes a test-file diff for human review.
It plugs into coding agents via MCP
- The MCP server lets your coding agent write, run, and debug tests.
- It checks the changed flow in a live session and saves the YAML.

Community Reactions

Codex is more precise and cheaper, commenters noted
- Quoted text says Mo has 84% precision versus 86% for Codex.
- Several commenters were confused that Mo costs $2.67 per bug versus $0.80 for Codex.
Commenters say a custom harness is enough
- Some said Claude with skills or a Playwright harness worked better for them.
- Others worried new models will absorb the edge of custom harnesses within months.
Pushback on cloud dependence and token billing
- One user found it slow because it runs in the cloud and expensive compared with subscriptions.
- Others wanted open source, self-hosting, and BYOA (bring your own subscription) as the default.
Silent false passes matter more than precision and recall, one commenter argued
- They said passing an already broken flow is the worst failure.
- They planted bugs on purpose to count misses, and allowing an 'I can't tell' answer mostly removed false passes.
Usability issues reported
- One user got stuck on auth setup with no onboarding.
- Another hit an error that their iOS app was incompatible with the hosted simulator.

How to Apply

Run a quick QA pass with your current agent
- Pick one feature you changed today and paste the prompt below into your coding agent.
- Compare the bugs and repro steps against what your existing tests missed.
Move core flows into plain-English YAML tests
- Write login or checkout flows as act and assert steps.
- If you use Momentic, commit the files so they go through code review.
Make the CLI exit code a required CI check
- Add the CLI to your CI job to run E2E on every PR and block merges on failure.
- Classify failures as app bugs, test flakiness, or environment issues.
Judge tools by cost and precision, not just bug count
- Compare cost per bug and precision alongside the number of bugs found.
- If you verify one changed feature per loop, check whether cost matters more than exploration breadth.

Try It

copy & run

# Paste into Claude Code or ChatGPT (a simplified version of Mo's approach: explore, then verify by reproduction)

You are a QA engineer. Explore the app below like a user and find bugs.
App URL: {{URL}}
Feature to test: {{feature description}}
Expected behavior: {{expected behavior}}

Steps:
1. Explore screens and user flows and list suspicious behavior.
2. Reproduce each candidate from scratch and mark only reproduced ones as 'confirmed bugs'.
3. List non-reproduced ones separately as 'cannot tell'. Do not guess pass or bug.

Output: for each confirmed bug, title, repro steps, expected result, actual result, severity.

Next: turn one confirmed bug into a Playwright test and add it to CI.

Terminology

E2E testA test that walks through a full user flow on a real screen from start to finish. It is like buying something in a shop yourself to check the whole process.

Auto-healA feature that lets a test find controls again when buttons move or get renamed. It is like a navigation app finding your friend's new address.

Recall and precisionRecall is the share of real bugs found, and precision is the share of reports that are real bugs. A wider net catches more but also brings in junk.

MCPA standard connector that lets coding agents call external tools. Think of it as a USB port for agents.

Flaky testA test that sometimes passes and sometimes fails with no code change.

Related Resources

Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app

Nov 2, 2026•wuweiweiwu•View Original

TL;DR Highlight

Mo, an AI QA agent that explores your app, found about 269 labeled bugs on QA Bench versus 86 for the next best.

Who Should Read

Web or mobile developers shipping fast with coding agents who struggle to write and maintain E2E tests

Core Mechanics

Mo explores your app from a single prompt
- Mo is an AI QA engineer that explores the whole app from a prompt.
- You review the findings, recordings, and repro steps.
Mo found far more bugs on QA Bench
- It ran 60 minutes on each of 20 areas across Gitea, Metabase, Twenty, and Saleor.
- Mo found about 269 labeled bugs, and the next best agent found about 86.
- The comparison was against Codex, Claude Code, and Devin.
A separate agent reproduces each bug
- A separate agent reproduces each functional bug before Mo reports it.
Tests live in your repo as YAML
- Steps and assertions are written in plain-English YAML.
- The CLI exit code can serve as a required CI check.
Tests repair themselves when the UI changes
- Auto-heal re-finds controls during a run.
- Permanent healing proposes a test-file diff for human review.
It plugs into coding agents via MCP
- The MCP server lets your coding agent write, run, and debug tests.
- It checks the changed flow in a live session and saves the YAML.

Community Reactions

Codex is more precise and cheaper, commenters noted
- Quoted text says Mo has 84% precision versus 86% for Codex.
- Several commenters were confused that Mo costs $2.67 per bug versus $0.80 for Codex.
Commenters say a custom harness is enough
- Some said Claude with skills or a Playwright harness worked better for them.
- Others worried new models will absorb the edge of custom harnesses within months.
Pushback on cloud dependence and token billing
- One user found it slow because it runs in the cloud and expensive compared with subscriptions.
- Others wanted open source, self-hosting, and BYOA (bring your own subscription) as the default.
Silent false passes matter more than precision and recall, one commenter argued
- They said passing an already broken flow is the worst failure.
- They planted bugs on purpose to count misses, and allowing an 'I can't tell' answer mostly removed false passes.
Usability issues reported
- One user got stuck on auth setup with no onboarding.
- Another hit an error that their iOS app was incompatible with the hosted simulator.

How to Apply

Run a quick QA pass with your current agent
- Pick one feature you changed today and paste the prompt below into your coding agent.
- Compare the bugs and repro steps against what your existing tests missed.
Move core flows into plain-English YAML tests
- Write login or checkout flows as act and assert steps.
- If you use Momentic, commit the files so they go through code review.
Make the CLI exit code a required CI check
- Add the CLI to your CI job to run E2E on every PR and block merges on failure.
- Classify failures as app bugs, test flakiness, or environment issues.
Judge tools by cost and precision, not just bug count
- Compare cost per bug and precision alongside the number of bugs found.
- If you verify one changed feature per loop, check whether cost matters more than exploration breadth.

Try It

copy & run

# Paste into Claude Code or ChatGPT (a simplified version of Mo's approach: explore, then verify by reproduction)

You are a QA engineer. Explore the app below like a user and find bugs.
App URL: {{URL}}
Feature to test: {{feature description}}
Expected behavior: {{expected behavior}}

Steps:
1. Explore screens and user flows and list suspicious behavior.
2. Reproduce each candidate from scratch and mark only reproduced ones as 'confirmed bugs'.
3. List non-reproduced ones separately as 'cannot tell'. Do not guess pass or bug.

Output: for each confirmed bug, title, repro steps, expected result, actual result, severity.

Next: turn one confirmed bug into a Playwright test and add it to CI.

Terminology

E2E testA test that walks through a full user flow on a real screen from start to finish. It is like buying something in a shop yourself to check the whole process.

Auto-healA feature that lets a test find controls again when buttons move or get renamed. It is like a navigation app finding your friend's new address.

Recall and precisionRecall is the share of real bugs found, and precision is the share of reports that are real bugs. A wider net catches more but also brings in junk.

MCPA standard connector that lets coding agents call external tools. Think of it as a USB port for agents.

Flaky testA test that sometimes passes and sometimes fails with no code change.

Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources