momentichenry

Nice! Let us know if we can help. Just to save you some time, I will say that in order to speed up data labelling, we pinned a commit a little bit behind latest main, since some of the bugs would already be documented in fixes or GitHub issues. So some of the bugs may have already been fixed!

Hey there! It's not strictly find a bug, as it's also getting the skill for the respective browser tools, but this is fair criticism. I think the current bench shows that if you don't have the time and money to build you're own harness, you're much better off using Mo. However, we do want to start including some of the harnesses that customers/users built before switching to Mo for comparison.

Have you built your own? Would love to see your results and add them to our site!

Thank you!

Currently it's only simulators and emulators, but we'll be rolling out real devices in the next few weeks.

Yes! We create a knowledge base entry so that Mo learns from its mistakes. This is especially helpful for stuff that's actually intended behavior, because at the end of the day, some things just come down to taste.

Hey there! This is coming soon! Although I will say, when we did an earlier run of this, Codex and Claude actually figured out how to look up the source code for the fixture in the public GitHub repo, and while it improved recall a bit, it also caused them to hallucinate more issues due to assuming that their reading of the source code was correct.

Mo is mostly looking for user-facing failures, or at least things that impact the user in some way. So if the app was super slow due to this, or was timing out, it would likely find it. However, if this is purely a backend cost issue, it currently wouldn't be likely to flag it.

Mo is able to inspect the network traffic, although I'm curious how many bugs you see that truly have zero observable impact, and if they do indeed have no impact, do they matter?

That's fair, and I think also something we've seen while creating this benchmark: it's incredibly difficult, if not impossible, to annotate every single bug in an app. This is actually sort of what we were trying to capture with severity-weighted score: the agent should get way more credit if it finds really bad bugs than if it finds a bunch of minor UI quirks, although this could definitely be improved upon when we release a v2 of the benchmark.

Hey there! Definitely good points here. I think some of these are covered in the benchmark, for example, false positives requiring investigation is equivalent to (1-precision), bugs that still escape into production is (1-recall), and real defects is recall.

Our goal with Mo is to make maintenance effort as low as possible while still maintaining quality. One thing we've been talking about separately is running a benchmark where all of the agents have virtually unlimited time (or something really high like 24h) and they need to find as many bugs as possible. Is that more similar to what you're looking for?

Hey there! Yes, using different models for performing the work and verifying is definitely something we've run into as well. In fact, within Mo we use one model for finding suspect bugs and a different one for checking its work to avoid this exact failure mode.

Mo doesn't need specs to run, but it does support PR review-mode, or just open ended exploration. If you do have specs you want it to execute, you can pass those to the agent as well.

Hey there! Thanks for watching our demo! Would love to run Mo on some other, public, benchmarks and share results. I'm curious how you're evaluating your framework's ability to find bugs vs just execute flows in an app? I also took a look at your site and it seems like you're product is more akin to Momentic regression tests, as opposed to Mo, as it depends on pre-written test cases rather than exploring itself. Maybe a better comparison would be our AI action?

I think it's true that our biggest gains vs other agents are in codebase-wide sweeps, because to your point, PR checks are limited in scope. However, there's still value in running Mo on those features because:

1. It should be pretty efficient using prior knowledge from other runs.

2. It helps expand the graph as your app evolves so that the DAG is always up to date.

Hey there! I'm not 100% sure how many came from Mo originally, but I can add a bit about our process:

- The candidate ground truth bugs are generated using runs from all agents

- Every candidate gets reviewed by the other agents to generate a recommendation (just makes human review easier)

- A human reviewer makes the final call (no candidates are discarded just because Mo couldn't reproduce)

And in terms of grading the final runs:

- A lot of the final runs still found a small percentage of novel bugs (5-10%), because it's extremely difficult to build a comprehensive list of all bugs, even after hundreds of runs.

- Nothing that didn't match was auto-discarded (read - it wasn't rejected because Mo couldn't repro).

- All unmatched findings were run through the same human review process as above.

- All the runs were regraded after human review to get a final score with zero unmatched findings.

One interesting thing we did notice during this is that Codex and Claude tend to converge on the same findings run over run (very few initially unmatched findings on final runs), whereas Mo tends to find a lot more novel bugs on ever run.