Big Idea

The argument in one line.

Running coding agents on cloud-hosted, always-on machines rather than a local laptop turns bug verification, QA, and security review into unattended background work, and the payoff comes from giving agents dedicated compute plus recorded proof of what they actually did.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already use Cursor, Devin, or a similar agentic coding tool and want a system for running that work unattended, overnight or alongside other tasks.
You maintain a shipped app with a bug tracker like Linear or GitHub Issues and want an agent to verify reports before you spend time acting on them.
You want AI-driven QA that tests real user flows and returns video proof, instead of trusting an agent's own claim that a change works.

SKIP IF…

You're not yet using an agentic coding tool — this assumes you already have Cursor, Devin, Claude Code, or similar in your workflow.
You need a beginner account-setup walkthrough — this is a workflow overview, not a click-by-click tutorial.

TL;DR

The full version, fast.

Cloud agents run on their own dedicated machine instead of the local laptop, so the work keeps going after the lid closes. The video covers four uses: reproducing and verifying a reported bug before touching code, fixing a bug and proving the fix with a recorded video of the agent testing it, running full QA sweeps that generate a numbered test plan and log every pass and fail, and scheduled automations that scan the codebase daily for vulnerabilities and bugs, opening a PR and pinging Slack when they find one. The setup behind it is a linked issue tracker (Linear via MCP), a configured environment with repo access, a skill that tells the agent to record its own testing, and a review loop (Greptile's /greploop) that keeps the agent iterating until feedback is resolved.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:35

01 · Cold open

Hook: half of agent usage is now cloud-based; overview of what's coming.

00:35 – 01:47

02 · What is a Cloud Agent

Defines a cloud agent as one running on its own dedicated computer, independent of the local laptop staying open.

01:47 – 05:34

03 · Way 1: Verifying reported bugs

Uses Cursor Cloud Agent to reproduce a Linear-reported bug in Pluto before trusting the report; the agent runs for ~2 hours using computer use to click through the app and confirm the stall.

05:34 – 06:46

04 · Sponsor: Depot CI

Ad read for Depot, a fast, programmable CI engine built for AI-agent-driven pipelines.

06:46 – 08:37

05 · Way 2: Fixing bugs with recorded proof

Devin fixes an AgentMail approval bug in Pluto and records itself testing the fix end-to-end as review evidence; also shows the agent catching its own mid-task failures.

08:37 – 10:44

06 · Way 3: Quality assurance

Agent builds and executes a full QA test plan for Pluto — 166 tests across 21 suites — logging pass, fail, partial, and blocked results.

10:44 – 12:56

07 · Way 4: Loops and automations

Scheduled cloud-agent automations scan the codebase daily for vulnerabilities and critical bugs, opening PRs and pinging Slack when they find issues.

12:56 – 15:07

08 · My setup

The four pieces behind all of it: Linear MCP integration, a configured cloud environment with repo access, a custom skill that tells agents to record video proof, and the /greploop code-review loop.

15:07 – 17:06

09 · Live demo and sign-off

Assigns a live Linear ticket to both Cursor and Devin cloud agents to show the hand-off in real time, closes with a personal note and a Bible verse.

Atomic Insights

Lines worth screenshotting.

Cloud agents run on their own dedicated computer, so the work keeps going even after the laptop is closed.
Reproducing a reported bug before fixing it took one agent run an hour and 54 minutes — done while the reporter worked their day job.
Recording the agent testing its own fix turns code review into watching proof instead of trusting a claim.
An AI-run QA sweep against one app produced 166 tests across 21 suites: 147 passed, 2 failed, 6 partial, 9 blocked by missing environment variables.
Scheduled agent automations can scan a codebase daily, open a PR with a fix, and message the team on Slack — no human has to kick off the run.
Giving an agent full reproduction context, not just a bug description, is what turns its fix into a permanent solution instead of a patch.
A cloud agent needs a configured environment — a forked repo, stored environment variables, and a running version of the app — before it can use computer use to test anything.
A code-review loop skill like Greptile's /greploop keeps an agent iterating on its own PR until every review comment is resolved, instead of stopping after one pass.
Even a passing-looking QA run can hide real gaps — 9 of 166 tests were blocked simply because environment variables weren't set up for the agent.

Takeaway

Cloud compute turns agent work into background labor, not a task you babysit.

WHAT TO LEARN

Verifying bugs, proving fixes, running QA, and scanning for vulnerabilities all become work an agent can do unattended once it has its own environment and a habit of recording proof.

02What is a Cloud Agent

A cloud agent runs on its own dedicated machine, so it doesn't depend on your laptop staying open or your local hardware.
Because the work isn't tied to your machine, a single agent run can keep going for hours while you do something else entirely.

03Way 1: Verifying reported bugs

Before spending time fixing a reported bug, have an agent attempt to reproduce it first — this confirms the report is real instead of a one-off fluke in someone's browser.
An agent with computer use can operate a live version of your app, click through the exact repro steps, and hand back a recording as proof the bug exists.
Giving the agent full reproduction context, not just the bug description, produces a fix that addresses the root cause instead of papering over the symptom.

05Way 2: Fixing bugs with recorded proof

Asking the agent to record itself testing its own fix, especially for UI changes, replaces trusting its claim with watching the proof.
Agents can catch their own failures mid-task, marking a step as failed rather than reporting false success.

06Way 3: Quality assurance

An agent can be told to plan and execute a full QA pass, every tab and every feature, producing a numbered test plan before it starts running tests.
A 166-test run across 21 suites came back with 147 passed, 2 failed, 6 partial, and 9 blocked purely by missing environment variables — a reminder that agent QA is only as complete as its setup.
Running QA this way avoids overheating a local machine on long test passes, since the work happens on the agent's own remote hardware.

07Way 4: Loops and automations

Scheduled automations can scan a codebase daily for vulnerabilities or bugs, open a PR with a fix, and notify the team directly, with no manual kickoff required.
These automations only run unattended because the work happens on a cloud agent — a local process still depends on the laptop staying open.

08My setup

A working cloud agent setup needs four things: an issue tracker connected via MCP, a configured environment with repo access, a skill that tells the agent to record its testing, and a review loop that keeps it iterating until feedback is resolved.
A code-review loop skill like /greploop won't let the agent stop after one pass — it keeps feeding review comments back in until all of them are addressed.

Glossary

Terms worth knowing.

Cloud agent: An AI coding agent that runs on its own dedicated remote machine rather than the user's local computer, so it keeps working independently of that machine staying open.
Cursor Cloud Agent: Cursor's hosted feature for running coding agents on cloud infrastructure, reachable from the web app, desktop app, or mobile.
Devin: Cognition's autonomous AI software engineer product, also offered as a cloud agent platform with its own environments and automations.
Environment (cloud agent): A configured workspace — a forked repo plus stored environment variables — that gives a cloud agent a running version of an app to test against.
/greploop: A skill that feeds a code-review agent's (Greptile) feedback back into the coding agent in a loop until every review comment is addressed.
Computer use: An agent capability that lets it control a virtual machine's screen — clicking, navigating, typing — to interact with an app the way a human tester would.
AgentMail: A service that gives an AI agent its own dedicated email inbox for sending and receiving messages as part of a workflow.

Resources

Things they pointed at.

02:27toolCursor Cloud Agent ↗

02:27toolDevin (Cognition)

01:41toolLinear

05:39toolDepot CI ↗

06:46toolAgentMail

14:10productPluto ↗

15:02toolGreptile / /greploop skill ↗

Quotables

Lines you could clip.

00:00

“50% of my AI agent usage now happens on the cloud. I've been fully cloud agent pilled.”

cold-open thesis stat, no setup needed→ TikTok hook↗ Tweet quote

04:32

“A bandage, but it's actually a permanent fix.”

tight contrast, stands alone→ IG reel cold open↗ Tweet quote

08:54

“I basically told the AI agent to QA my entire app.”

concrete, surprising claim→ TikTok hook↗ Tweet quote

15:09

“It just this makes it feel like a video game.”

candid, relatable aside→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogy

00:0050% of my AI agent usage now happens on the cloud. I've been fully cloud agent pilled. I can't stop, and it's without reason.

00:09My productivity has increased, and I'm gonna explain all that to you in today's video. We're gonna talk about what a Cloud Agent is, the four different ways that I use it, how I have my Cloud Agent environment set up, and finally, you don't wanna miss this part. I'm gonna show you a full run from the beginning to the end.

00:24You don't wanna miss today's video. Sit back, relax, get your popcorn. Let's get started.

00:29So first things first, we gotta talk about cloud agents. What is a cloud agent? A cloud agent is an agent that runs on a cloud.

00:36But in all seriousness, there's a main benefit to this. Right? And the main benefit is that cloud agents have their own dedicated computer.

00:44And with this computer, with its own hardware, there's a lot of cool things that you can do. First things first is avoiding this. I don't know if you've experienced this, but, you know, back in my day when we were writing code, after we were done, the laptop stayed closed.

00:58But now with agents, we can send off a couple prompts. You can have a couple threads running. You're gonna keep that laptop open because if it closes, all the processes are shutting down.

01:06And this is how most people are working now. This is the one main thing you have to understand about cloud agents because it run-in its own dedicated computer. The agent can run for however long it needs to run until that task is done.

01:19It is not dependent on me or my machine. Side note, this is why I'm building a home lab so I can have my own cloud, but make sure you're subscribed so you can watch that video. And please, please, please do subscribe.

01:28We're almost at a 100 k subscribers. But going back to my original point, that is one of the main benefits. It has its own computer.

01:34It runs on the cloud. This thing can go on and on. But there's some other use cases that people aren't talking about that I'm gonna be sharing with you today that has made me not only productive, but even the quality of applications that I'm creating to be fantastical.

01:47So ways that I use it. First is verifying reported bugs, and I'm just gonna show you a live example instead of telling. So if you watch my videos, you know that I use linear for issue tracking, and Muso, my business partner, tagged me in this ticket.

02:03This is a bug in one of our apps called Pluto that we've been building. He says it says here agent run stalls in streaming state and inspector reports stalled run.

02:12So the problem, the user report, and I even have a video on a screenshot. And listen.

02:16I trust Musso. He's my homie. Look at this handsome face.

02:19You make sure you follow him on Twitter. But I need to make sure that this wasn't just one isolated incident that, you know, happened to him or maybe his browser was funky or something happened or because, you know, he did use AI to generate. So and I need to 100% make sure that this bug is real.

02:35Now there are two Cloud Agent platforms that I love, and I'm using both of them. And that's Cursor Cloud Agent and Devon Cloud Agent.

02:43You all remember Devon, the software engineer? Devon's Cloud Agent platform is great. Cursor's Cloud Agent platform is great, and I'm gonna show you both.

02:49But in this particular example, I'm going to be showing you Cursor's Cloud Agent of Arm. You could see I use this a lot. So I went on cursor.com.

02:57Again, if you wanna use the Cloud Agent platform, you have to go on the website. You can also do it on the desktop app. I prefer the web app, and you they also have mobile apps which are great too so I can work from my phone.

03:06But I basically said pick up the issue P l o 230, which is the ticket we just read from the linear issue, but first test of the bug is actually legitimate or not. Try to reproduce it. It worked for an hour and fifty four minutes.

03:20It's not to say now here's what you might say, oh, but what if I know this bug to be true? I genuinely believe giving the agent context and reproducing the bug allows for the agent to fully understand the issue, to fully diagnose the problem, and to fully create a solution that fixes this.

03:37So it worked for almost two hours. You know what I was doing for these two hours? I was working my day job.

03:43I was working. I was grinding. I'm building.

03:46You know what I'm saying? Once it was done, look what it produced for me. Look what it gave me.

03:50It gave me a video. This is the video of the Cloud Agent using its computer.

03:57It has computer use. It has a Linux machine, and it's reproducing the error.

04:03It's literally using my app live, like, Well, not live right now. This is recording.

04:09But you could see it's clicking around. It's changing tabs and that it confirmed that it's stuck.

04:17It stalled. So you could see I got a recording of the issue. So now it knows why it happens.

04:24This is process run marks, chat runs, chat messages, streaming before dispatching, blah blah blah blah blah blah blah blah blah. It knows exactly what the issue is.

04:33Now I can tell the agent, please fix it. And it has because it has the context of how the bug happened, what the issues of the bug are, it's going to do a better job at presenting a fix that's not only just like, you know, what do they call it? What's the, uh, saying?

04:48A bandage, but it's actually a permanent fix. Another example here, I said pick up the issue p l o two two eight from linear, but first test if the bug is actually legitimate or not, then try to reproduce it, and we see this time it shared a screenshot of the bug reproduced.

05:03So that's way number one. That's verifying reported bugs. And number two is actually fixing bugs, but check this.

05:09I want proof of the bug fix with a recording of the agent actually testing it out, especially if it's a UI centered type of change. Like, let's say there was a back end issue and the chat was broken. I want the agent to test the chat, send a message, record that, and show that to me.

05:24This is part of the review process or my human review process where I needed the agent to give me a recording. Come here. Let me ask you a question real quick.

05:31When was the last time your CI pipeline finished before you lost interest?

05:39Exactly. And this tells me you're using something slow and ugly like GitHub actions and it's time to upgrade. Meet Depo CI.

05:47This isn't just another CI tool by the way. It's a CI engine that's incredibly fast, fully programmable, and genuinely built for the AI era.

05:55Meaning, you can literally give your coding agent access to Deebo CLI or API and it will set up everything for you. I'm talking you can spin a pipeline. Everything is configurable and you can get things running with just a few commands.

06:07And listen, don't just take my word. Companies like Posthog and Wistia are using Deebo CI as we speak. And once you switch, waiting for CI to take forever becomes a thing of the past.

06:16Now one thing you might think is something as good as this must cost an arm and a leg. I'm here to tell you cost $0 and $0.00 $0.01 cent per second of compute. That means if you wanted to spend $3, you'd be getting five hundred minutes worth of compute.

06:29So the question isn't how much does Depot CI cost? The question is how much is slow GitHub actions costing you? Really think about it.

06:36Life's too short to spend time with GitHub actions. Get on to Depot CI. The link is in the description down below.

06:41Let's get back to the video. Now here's an example of Devin fixing a bug. Now if I go to Pluto, you'll see under the more tab, there's an email tab and users can give their agents its own dedicated email, and we're using agent mail as a service provider.

06:55Now if I go back to Pluto, you'll see here I had an agent mail issue where, like, emails are just not being approved. Like, when an email was drafted and it asked the user to approve it, if you clicked approve, it just wasn't working. So told the agent the issue, and the agent proposed the fix, made the fix, created a PR, but with the PR, recorded a video of it the agent itself testing it.

07:19So check this out. So the agent clicks on new chat, says send an email from agent mail inbox, and then that fires off, clicks approve, and then it checks the trace, and an email was sent.

07:32So reviewing is a lot easier for me now because now not only can I assume that the agent did the right thing, but I almost am certain because I see it live? It's being tested here live.

07:45Now you might think, how out of the agent just hallucinates and thinks that it works when it doesn't? There are many cases I can show you where if I scroll up here, it was doing a test for a feature and it failed in some places. And you can see Devin here marks it as failure, marks it as a bug.

08:01And then another time here you see upload okay, but index failed. Platform OpenAI embeddings quota. Uh, Yeah.

08:08The credits. The credits ran out. That was more of a money issue, but you see the agent caught it.

08:12Right? So with Cloud Agents, I can have this happening in the background while I'm doing other things. This sort of ties in with the previous video I did.

08:20Make sure to check it where I talk about loops. Loops can only be successful and uninterrupted if I'm using something like a Cloud Agent.

08:27Number three, quality assurance. What do I mean by this? So back in the day when I was working at software developer job at a corporate company where, you know, you would get Jira tickets, you would propose fixes, you would send PRs, There was someone known as QA quality assurance and their job was to test out the feature and borderline try to break it to see if it works.

08:47Now I wanted to do this test the other day and it was insane that the fact that I can have AI do my QA. So I basically told the AI agent to QA my entire app, and I'll show you the exact prompts.

09:03Look. I first said I want you to create a testing plan where you go through every single tab, every single feature, every single thing that one can possibly do with Pluto. I want you to test it, document it, record a video.

09:14If it passes, it passes. If it fails, we note this down, then we're going to fix it. But I want us to just first test every single feature.

09:21Make sure everything works. Let's plan it out. I think first things first, let's create a plan on what we're going to test and how many tests we're going to need.

09:28So it created a plan, 166 tests across 21 suites. What I want to show you is the test. So we're seeing it test the onboarding flow.

09:38It's trying to enter a non email, navigation, like, there's so many videos in here.

09:44Right? To check-in mobile, like, all these different things, like, it went on and on and on for hours. And then I got a report, a 147 passed, two failed, six were partial, nine were blocked by an environment variables, meaning I didn't set up the Cloud Agent with all the environment variables it needs.

10:00I'll show you that later in the video. And then two were skipped. One of them was retired feature and the other one was a guardrail.

10:06But you can see how instead of me having to manually test every single thing, if I have a cloud agent, I could tell it, go test these eight features, prove that they work, and record a video for all of them or maybe record a video of the features that don't work out so I have context as to what's broken and how it can be fixed.

10:24All of this stuff is possible when you're using Cloud Agents because if you're running this on your machine, I have a beefy MacBook. The fan will start fanning.

10:32Things will start getting loud. Things will start getting slow. Things like this, you need a cloud agent.

10:36You can also do this with Devon as well. And finally, loops and automations. Now I'm not gonna talk about the loops that I use.

10:42Please watch the video before. I talk about it in detail. But automations is something cool that a cursor in particular has.

10:49If I go to the automations tab, there are a couple predefined ones, example ones here that they have, but there's two in particular that I have that are pretty interesting. This first one is basically, it scans the code base for any vulnerabilities. It runs at 1PM Eastern Standard Time every single day, and there's this detailed instruction that basically tells it to check every single part of the app and scan for vulnerabilities.

11:14And when it finds one, it's going to create a PR with a code fix, and it's gonna hit me up on Slack. I have Slack triggered. I'm gonna get a message on Slack and the Slack message is basically gonna be this was the issue.

11:25This is the PR to fix it. Please review and merge. This happens every single day.

11:30Another example I have here is one for finding critical bugs. It's the same one as the security one, but this one is just for bugs. Right?

11:37So this scans the code base every single morning, 7AM. Now you might ask why 7AM and why 1PM for that. Just picked random times.

11:45It doesn't really matter. And again, the agent instructions basically tell it to review the code base and to find bugs. And once it's found something, it's gonna create a PR with a fix, shoot me a message on Slack.

11:56All of this is thank you to cloud agents because when my laptop is closed, this type of stuff can't happen. Now how does my setup work? There's four particular things that I cannot live without if I'm using Cloud Agents.

12:08First, I'm linking the linear MCP. Now I don't need to show you how to do this, but for cursor, it's literally as simple as going to settings and then I think it's integrations.

12:20Right? Integrations. And then you just connect linear.

12:23Right? Same thing with Devin. If you literally go to settings and then you go to connections and then I have linear connected.

12:30So linear becomes a place where we track feature requests, bugs, and all that stuff. And what's cool about the MCP integration is, let's say, for example, there's this bug that Musso caught, and I wanted to get the cursor agent on the cloud agent.

12:44All I literally have to do is assign the agent, and it's gonna start getting torque. I'll show you that at the very end. Remember, I promise to show you a full run.

12:51Right? So that's number one. I have a place where I track issues and bugs.

12:57And for me, it's linear. For you, could be GitHub issues. I don't know why you would do that to yourself.

13:02It could be Trello. It could be Notion. Whatever it is that you like to use, use.

13:06I really like linear. The UI speaks to me. Number two is setting up your environment.

13:11And basically what that means is both, again, cursor and dev and do this. If I go in the settings for cursor again and I click on cloud agents, you have this environments tab. I can click new here and, basically, what this does is it allows me to create an environment.

13:25An environment is basically in your computer, the agent will fork the repo, and then it will ask you for environment variables. You'll give the agent the environment variables are stored securely in the platform.

13:36The Cloud Agent platform does it for you. And now every time I call the Cloud Agent, it has access to a running version of your code base. Right?

13:45So if I go to, let's say, whichever tab, let me go to the test tab right here and I click on desktop, you're going to see on the computer a running version of the app.

13:58And the reason why this running version of the app exists is because I set up an environment. It's the same thing with Devin. Right?

14:05With Devin, all I do is click on settings, go to environment, and then I click on add here, and then I add a repository, do the same thing.

14:13So every time an agent is running, I spin up a new thread. It has full access to a running computer that has my app running. This makes it so that the agent can use computer use and test my app after shipping the change or reproduce the bug.

14:29And finally, the last thing that I use, I mentioned this in many videos is slash greploop. If you aren't familiar with greploop, greploop is a skill by the folks over at Greptile, which is one of my favorite code review agents. And what greploop does is it basically runs the Greptile code review feedback back into the agent.

14:48And until the agent addresses all the feedback that the code review agent gave it, it's not going to stop running. It's gonna keep looping again and again and again for all the loop stuff, please watch the video I did the other day, like the most recent video other than this very one that you're watching. Now finally, let me show you a Cloud Agent in action.

15:09I'm gonna go to linear and then I'm not even gonna read this issue. I'm going to assign cursor and then you could see that nice little animation. See, this is why I love linear.

15:19Like, it just this makes it feel like a video game. It feels great. Now if I go back to cursor, you can see here that my Cloud Agent has spun up.

15:28It has context from the ticket. It's going to start dissecting the issue, and it's going to start working on a fix. Now this might take some time.

15:36This might take twenty minutes, thirty minutes, an hour, whatever it is. I'll show you another example where I'm going to fire off Devin. For Devin, I'll do this one as well.

15:45All I have to do again is assign the Devin agent. And if I go back to Devin, you can see here that the Devin cloud agent starts to work.

15:54Now one thing I forgot to share is if you go on my GitHub profile, I'll link this in the description. If you go to my skills repository, I just added this evidence driven testing skill.

16:05This is basically the skill that tells your agent to record a video, to verify the changes, to do a test. All that stuff that I just shared with you, I've turned to a skill. The link will be in the description down below.

16:17And finally, this is sort of the best part of Cloud Agent's like, yeah, I just it's just doing its thing and I'm kinda free. I think I'll pick up my bible and and just read.

16:27Yeah. That's it. I just gotta let it work.

16:28In fact, why don't we end off this video with a bible verse? Psalms chapter one verse one says, all the joys of those who do not not follow the advice of the wicked or stand around with sinners or join with mockers, but they delight in the law of the Lord. Meditating on it day and night, they are like trees planted along the riverbank bearing fruit each season, their leaves never wither.

16:50They prosper in all that they do. That's pretty much it for this video, guys. Make sure to like, comment, subscribe, hit the notification bell.

16:57I hope you understand Cloud Agents. If you have any questions, let me know down below. You've been awesome.

17:02My name is Ross. I'll see you in the next video. Peace.

The Hook

The bait, then the rug-pull.

Half of this creator's AI agent usage now happens on the cloud instead of his own laptop — and he says it's changed not just his output, but the quality of what he ships. What follows is a tour of four concrete ways he puts cloud agents to work, and the setup behind all of them.

Frameworks

Named ideas worth stealing.

01:48list

4 Ways I Use Cloud Agents

Verifying reported bugs
Fix bugs and prove fix via recording
Quality Assurance
Loops / Automations

The four recurring use cases the video organizes around.

Steal forany team workflow that wants agent work to run unattended

12:01list

My Setup

Linking Linear MCP
Setting up environment
Telling agent to record video of bug fix (custom skill)
/greploop

The four pieces of infrastructure that make the four use cases possible.

Steal forstanding up your own cloud-agent workflow

CTA Breakdown

How they asked for the click.

VERBAL ASK

05:39product

“Meet Depo(t) CI... The link is in the description down below.”

Direct sponsor read framed as a pain-point contrast against slow GitHub Actions, backed by a specific price ($0.01/sec of compute) and a single link-in-description CTA.

MENTIONED ON CAMERA

02:27toolCursor Cloud Agent ↗

05:39toolDepot CI ↗

14:10productPluto ↗

15:02toolGreptile / /greploop skill ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.