E2E tests are now quick to write due to LLMs, and are then deterministic AND cheap to run. How would this compare to the token costs of running an agent the whole time for each test? How do you make sure results stay stable regardless of the nondeterministic nature? Do customers still need to create test cases - any way to import from test case management system - based on which they could have already generate e2e tests locally?
Unfortunately from our experience tests don’t scale as well as code. First of all, static tests are very brittle: you rely on selectors, need wait times, and can’t really test a lot of dynamic content (think AI chats/interactions). Then it’s all the infrastructure around it: solving captchas, handling auth, handling email OTP (each of our agents has access to its own inbox) and handling video recording and screenshots.
To ensure stable results we do a lot of harness engineering, where we inject trajectories of previous tests to ensure the stability and also the split into smaller steps helps to prevent context overload and decision fatigue.
Regarding test case management, our customers have used our CLI to migrate their existing test cases from whatever system they were using before.
We support both web and mobile, which is what a lot of companies prefer, just one agent for both. Also, I'm pretty sure Revyl relies only on vision models, which tend to be slower. We built the platform around a hybrid approach that combines vision and accessibility APIs, which is much faster.
Would love to hear your feedback after you try it out!
"Traditional E2E tests are slow to set up and expensive to maintain." I don't really understand this. If I'm already using Opus to write the code, surely it would know best what E2E tests to write to be able to verify its own output? This seems like an unnecessary external step.
Unfortunately from our experience tests don’t scale as well as code. First of all static tests are very brittle, you rely on selectors, need wait times and can’t really test a lot of dynamic content (think AI chats/interactions). Then it’s all the infrastructure around it: solving captchas, handling auth, handling email OTP (each of our agents has access to its own inbox) and handling video recording and screenshots. So with the traditional testing approach you end up mocking a lot of services. I highly recommend you to give it a try!
Have you been able to nail down a loop where your tool can take an open pr, guess the code path and do some testing?
We use cypress heavily for our core flows which has a similar ai prompt thing but it’s not quite ad hoc enough for smaller fixes which is where the bottleneck still comes in for us.
Yes! We spent quite a lot of time on this, and we are currently creating a test plan based on PR changes and sending an agent to verify it. We have some customers who are only using this feature.
We use agents to navigate the app, making real-time decisions based on its state. I prefer to compare it more to a manual QA engineer than to static e2e tests. We spent a lot of time on the harness to make sure the results are reliable. This allows you to assert on dynamic content like AI-generated content. We also support validation of email flows since the agent can read its own email.
Fable (rip) is absurdly good at this, great time to build a product around it, you definitely need the harness, but it feels like it just turned the corner to be able to do really in depth and edge case work.
Do you handle heterogenous environments and network connectivity simulation as well? I am working on a mobile app and occasionally having users just lose a request or two can put the state machine into unusual modes.
I feel like new AI model releases will only allow our agents to do more in-depth testing; the space still has a lot of room to grow. Quality assurance is way more complicated than just clicking around a UI.
Regarding the other question: not yet. For now, we have Chromium, iOS, and Android (latest versions of each), but we are working on adding more. Regarding network connectivity, it's coming soon (I have an open PR).
Does it work of mobile native applications or expo apps that have native modules?
Pricing question, the usage on the plans seems low considering in the demo you said that you have 25 tests per pr which would mean you get only 10 PRs per month on the hobby plan?
Yes, it works for any framework. We just get the built native binary and run it in the cloud.
Regarding pricing, the self serve options are currently only for lower usage. We will add more plans further down the line. Currently the most popular one is the startup plan. If you need more usage I’m happy to discuss it on a call!
E2E tests are now quick to write due to LLMs, and are then deterministic AND cheap to run. How would this compare to the token costs of running an agent the whole time for each test? How do you make sure results stay stable regardless of the nondeterministic nature? Do customers still need to create test cases - any way to import from test case management system - based on which they could have already generate e2e tests locally?
Unfortunately from our experience tests don’t scale as well as code. First of all, static tests are very brittle: you rely on selectors, need wait times, and can’t really test a lot of dynamic content (think AI chats/interactions). Then it’s all the infrastructure around it: solving captchas, handling auth, handling email OTP (each of our agents has access to its own inbox) and handling video recording and screenshots.
To ensure stable results we do a lot of harness engineering, where we inject trajectories of previous tests to ensure the stability and also the split into smaller steps helps to prevent context overload and decision fatigue.
Regarding test case management, our customers have used our CLI to migrate their existing test cases from whatever system they were using before.
I'm curious how your mobile testing compares to https://revyl.com
I've been experimenting with Revyl and it's really nice. I think this agent-driven testing is the future.
We support both web and mobile, which is what a lot of companies prefer, just one agent for both. Also, I'm pretty sure Revyl relies only on vision models, which tend to be slower. We built the platform around a hybrid approach that combines vision and accessibility APIs, which is much faster.
Would love to hear your feedback after you try it out!
Love your approach to product. It feels like TesterArmy will become the "Vercel for testing". Refreshing stuff!
Thank you! That's the goal
"Traditional E2E tests are slow to set up and expensive to maintain." I don't really understand this. If I'm already using Opus to write the code, surely it would know best what E2E tests to write to be able to verify its own output? This seems like an unnecessary external step.
Unfortunately from our experience tests don’t scale as well as code. First of all static tests are very brittle, you rely on selectors, need wait times and can’t really test a lot of dynamic content (think AI chats/interactions). Then it’s all the infrastructure around it: solving captchas, handling auth, handling email OTP (each of our agents has access to its own inbox) and handling video recording and screenshots. So with the traditional testing approach you end up mocking a lot of services. I highly recommend you to give it a try!
Have you been able to nail down a loop where your tool can take an open pr, guess the code path and do some testing?
We use cypress heavily for our core flows which has a similar ai prompt thing but it’s not quite ad hoc enough for smaller fixes which is where the bottleneck still comes in for us.
Yes! We spent quite a lot of time on this, and we are currently creating a test plan based on PR changes and sending an agent to verify it. We have some customers who are only using this feature.
Seems interesting, but I wonder about this
> Traditional E2E tests are slow to set up and expensive to maintain.
Isn't this just using agents to create e2e tests or is there some better new approach I'm missing?
We use agents to navigate the app, making real-time decisions based on its state. I prefer to compare it more to a manual QA engineer than to static e2e tests. We spent a lot of time on the harness to make sure the results are reliable. This allows you to assert on dynamic content like AI-generated content. We also support validation of email flows since the agent can read its own email.
Fable (rip) is absurdly good at this, great time to build a product around it, you definitely need the harness, but it feels like it just turned the corner to be able to do really in depth and edge case work.
Do you handle heterogenous environments and network connectivity simulation as well? I am working on a mobile app and occasionally having users just lose a request or two can put the state machine into unusual modes.
I feel like new AI model releases will only allow our agents to do more in-depth testing; the space still has a lot of room to grow. Quality assurance is way more complicated than just clicking around a UI.
Regarding the other question: not yet. For now, we have Chromium, iOS, and Android (latest versions of each), but we are working on adding more. Regarding network connectivity, it's coming soon (I have an open PR).
Does it work of mobile native applications or expo apps that have native modules?
Pricing question, the usage on the plans seems low considering in the demo you said that you have 25 tests per pr which would mean you get only 10 PRs per month on the hobby plan?
Yes, it works for any framework. We just get the built native binary and run it in the cloud.
Regarding pricing, the self serve options are currently only for lower usage. We will add more plans further down the line. Currently the most popular one is the startup plan. If you need more usage I’m happy to discuss it on a call!
Congratulations on launch, I’ve been tracking your progress since you’ve been accepted for spring batch.
Always happy to see cool products from Poland! :)
Thank you!
.army?
We are thinking whether to change this.. We also have testerarmy.com/.ai
Change it now to .com or get stuck there for years, suffering anti spam filters, potential renewal problems and more in the meantime.