The Local Dev Env is Dead

Thorsten: Most developers that I talk to underestimate how useless the dev env... the local dev env now is like look, what was always the reason why you would never use cloud IDEs? It's "oh it's not my editor, I cannot use my terminal, it doesn't have my dot files, it doesn't do this, it doesn't... blah blah blah my shortcuts", like all of that stuff. But now it turns out I don't care about the text editor shortcuts anymore.

Thorsten: I don't care about my shell tab completion anymore. I don't care about the language server anymore. I send prompts to my agent and then review the work.

Thorsten: If you start from first principles, a lot of people miss That a lot that the reasons why they think local dev is is required they're obsolete. Welcome back, Raising an Agent Season 2. It's been a while. It's good to be back.

Quinn: It's good to be back. It's been a while since we got on and talked through this stuff. We've done a lot. We went heads down.

Quinn: We rebuilt Amp. We came out with this thing in February saying, "the coding agent is dead". And then now we're back shipping constantly.

Quinn: Now we didn't do nothing. We were heads down building and we'll talk about you know what we built, what made us radicalized and wanna go in this heads down mode. Um and then we're also just, you know, growing a lot.

Quinn: Because it turns out in this funny world, when we think that what we have, the current tech tree is obsolete, that actually is when it's just starting to hit the mass market and the growth is really strong. So it's really weird. This market is really weird.

Thorsten: Yeah. So to recap, in February we said the coding agent is dead, and what we meant was that the coding agent, as most people have used it in the prior six months, where you have it in the sidebar of the editor.

Thorsten: Or like one terminal pane, maybe two, three, but you manage them and you actually put context in and out and you have like your custom splash commands and all of this. Um we thought that's dead. Like we we the And I I don't know.

Thorsten: I think I put it in the writing, right? Like uh "it might not be dead yet, but the bullet has has left the chamber". Like it yes, we still use it locally, but it's kinda obvious that this uh the thing in the sidebar, at least that's completely dead.

Thorsten: I don't think anybody still does that really. I maybe there's Copilot users that disagree.

Quinn: I mean the the thing is probably the "majority of developers" still is there, but they're not gonna be there for long.

Thorsten: Yes, yeah, I think so. I think I think yeah, and and I mean go back to the first episode here, like if this is not about the majority of the you know, like we're not talking about like the 90% of what's coming, but we've been pretty pretty good in you know being somewhere that where the majority then will catch up. So yeah, we've done we said like coding agent is dead and from our perspective, the next logical step would be one layer up or one level up where you don't have to manage a single agent and and and and worry about like you know context stuff and so on so we went one layer up and thought okay the models are now so good, what does the next layer look like?

Thorsten: And then we rewrote Amp based on this. And we came out in May with Amp Neo. That was our release or codename for the new version of Amp.

Thorsten: It had compaction, which I'll get to in a second. It is remote controllable and still has compaction, sorry. It is everywhere controllable from the web, from your phone, from your TUI, and It has a plugin API so you can customize it because it turns out agents are really good at writing plugins and a lot of stuff that previously went into the harness, an agent can now whip up.

Thorsten: Just just to drive the point home: where we were and where we are right now. I think Lewis published his post in December maybe last year about "200k context window is enough".

Thorsten: I think he called I think it was in November. I'm relatively sure. And we had our first version of what came to be known as Amp Neo in March, I think.

Thorsten: Like we were ourselves we were pretty fast on using the new thing. Took us a while to ship it, which we can get to in another episode. But in that new version we had compaction.

Thorsten: Because we realized the models are now good enough at compaction. They do they don't lose track as before and they don't go off the rails after compaction, whatnot. And I think it was a week of using it where people in our internal Slack went, "I can never go back".

Thorsten: Like this is "I'm never gonna manage context manually again".

Quinn: Yeah. And the funny thing is there's a lot of people that said, "oh, I love an Amp because it lets me manage my context better than anything else". "I love that you made this handoff thing and you should really go focus more on that".

Quinn: And it's so dangerous when you're building an agent to get tempted to do those things, the things that you're known for being good at.

Thorsten: Yeah.

Quinn: But no, we went.

Quinn: deep on compaction. And yeah, I don't think anyone who's used compaction, I don't we don't hear anyone saying "I'd rather go back to the old way".

Thorsten: And I think I mean we've been pretty good with deleting features and you know removing stuff and...

Quinn: We're known for that.

Thorsten: Yeah, going against the wishes of our users, maybe at some point. And I think that was one of the most brutal ones were... because we were known for handoff and like some of the primitives that we offered, right?

Thorsten: Like handoff was I think a good feature people loved it and we have like the relationships between threads and Lewis built the threat map in the TUI amazing thing. It's beautiful. And we said that's that's over.

Thorsten: And and and it was really hard. You know, as somebody who really loves removing features, I think that was really hard. But it's just now looking back, obviously the right choice.

Thorsten: And it's just I think that's a good example of what I mean that coding agent is dead. This like handheld managing of a single instance of an agent, and we got to move up one layer. And now look at us.

Thorsten: I don't care about the context window size anymore. I don't care about you know, handoffs and whatnot. Like everything is just one level higher.

Thorsten: And that was the goal with Neo. So we shipped Neo, but underneath the architecture is completely different now. The architecture is we can do another episode on this, but basically it allows us to run agents anywhere you want on any machine you want and control them from everywhere, from the same interface or from your TUI, from the phone, from the web.

Thorsten: You can spawn an agent on your Raspberry Pi, on your dev box, in orbs, which we'll get to in a second, and control them from everywhere. And I think that is that's the future. Like I think it's less about you sitting there on your machine in your dev setup in this shell in this directory and you watching multiple checkouts, which already sounds insane when you think about it.

Thorsten: The new the new way of working is you're just worried about what do you want to get done and and how to verify whether it's correct or not. But it's not like the managing of things. Yeah.

Quinn: Yeah. And you know, if you think about it, a year ago, Amp and everyone who was joining us using Amp back then, we were all experiencing what it was like to not be restricted in your token usage to have an agent go and take a lot of steps on your behalf compared to what everyone else was using for, which was chat, which one response. I mean this old history by now, but it felt so free and you could do so much more when an agent could just keep on using the tokens and it sounded crazy at the time, obviously there's so many different kinds of workflows that needed to change.

Quinn: People were talking about needing to approve every single edit and they would actually like nitpick the agent on its small edits. Now we know that agents can just steamroll past those things. But now I think there's this similar kind of moment where if you have the ability to run agents anywhere and to run them in these little isolated, independent lightweight environments, orbs, we call them, and that name seems to have stuck, then it opens up so much more parallelism, and it means that all these things that you considered so important you don't care about anymore and the most interesting one maybe that we've seen so far is: we don't care about local dev anymore. We think that local dev used to be, you know, up there with CI and prod.

Quinn: You had to have really good local dev.

Thorsten: Yeah.

Quinn: And we don't think it matters anymore on the frontier.

Thorsten: I think it's... the bullet has left the chamber. Like, yes, I do stuff locally, right? But at this point it's a question of time, I would say.

Quinn: So what do you do locally that you cannot yet do in an orb in the cloud?

Thorsten: The only thing I do locally is stuff that needs some credentials and whatnot.

Quinn: Some credentials that we have not yet been able to safely get into an orb.

Thorsten: Yeah like production database access, something like that. Yeah might be... we need to figure this out still. But yeah. Yeah, I I just to set the scene, we jumped in this real quick.

Thorsten: So in the last six weeks, we launched, you know Amp Neo, we launched a new web interface where you can manage all of your threads from anywhere, web, phone, whatever. And then three weeks ago, I think three weeks ago, we launched agents in orbs, and orbs is our name for what others might call uh Fast to start up, quick, ephemeral, isolated sandbox. So what we have in our model is: you start a new thread and you get a new orb attached to it.

Thorsten: And it's not like you have a single orb or five orbs or three orbs and you pick and start threads in them. No, you start a new thread from ampcode.com, you select your project, your repo, it boots up an orb.

Thorsten: It should be relatively quick. Your repo gets cloned. Then snapshotted, and then you run your agent in this sandbox.

Thorsten: And then in the web, you can see the files in it. You can add the terminal, you can use a terminal, and you can talk to the agent. Um, and I think It others also do this a little bit, but for us, what I think is interesting is we've always had to thread as something that you can share

Thorsten: in Amp like you can share the conversations with your teammates and whatnot and now like a thread ID or thread URL is also tied to the orb. It's tied to the the computation that created the changes in this. It's tied to the context, and you can share basically the whole thing, which we'll get out in the next few days, which we're working on.

Thorsten: But that is the thing. And to to to recap, I don't think we made this is maybe the redemption year, but I don't think I did a great job in the blog post of explaining how things change for us. And I want you to explain.

Thorsten: I put in the blog post that I went into this thinking that orbs are kind of like my remote, sorry, my local dev, but remote. Where "Oh, I have copies of my local dev and I can remote control them". You know, like the use case of if you go to lunch and you know whatever on your machine we want to start a thread or something, except now you replicate your local dev remotely and that's kinda the answer to this. But what we found going in is: if you have ephemeral orbs or orbs that are tied to a thread, it changes how you approach a lot of stuff. It changes even you do things that you wouldn't do on your local dev, which is I sometimes start orbs to fix CI and it's like an async functions.

Thorsten: I tell them "fix this and then push to main". And now obviously we've this is coming soon too. It's automated.

Thorsten: I don't even have to do this anymore. Or I start explorative things overnight. And then come back to it the next day and go through them one by one.

Thorsten: I don't have to worry about any state. I don't have to worry about any checkouts. Like I have six local checkouts.

Thorsten: If I had them locally, now there were still conflicts and you still had to worry about ports and whatnot. But now I don't know, like it's a different way of working. And yeah, how how has it changed for you?

Quinn: I have agents going on way longer explorations. I'll have them go and test things manually. No like "write unit tests and then run them", but "go and test things manually in every possible permutation".

Quinn: Like I made it so in the Amp CLI, after you started a thread, if you hover over the mode name. It shows a little, you know, image of uh a hover of what uh mode you've selected. And I had it try that for all the different variants that I knew.

Quinn: And I had it tried in different terminals, uh like you know, through Tmux and Knot and That's the kind of thing where I would never do that manually. And if it were on my laptop, I don't want to tie up my laptop for like four hours doing that and maybe I could get our local devs so that it is isolated if I have different work trees or different checkouts and I have that I actually had this really nice setup that lasted for about two months where I had um amp1, 2, 3, 4, 5, and there were different checkouts of our Amp repo. In Ghosty, the terminal background color would be different so I could distinguish between them, each of them.

Quinn: I had five tabs, but diff, you know, all that. And the problem is, even if you try so hard to remove uh dependencies so that things are totally isolated. You never totally quite get it.

Quinn: It's like, you know, the TLS certificate or like, you know, host names or someone else adds a service. And I just wouldn't want to tie up my laptop on these goose chases. But now I don't even have to think about it.

Quinn: There's no thought in my mind. And this just reminds me of back in Amp when uh, you know, in the right when we had launched it, when so many other people were in this mindset of "I need to review everything the agent does". And if you can just get that out of your mind and just figure out, okay, the world has changed.

Quinn: We need to plan for the agent is gonna be running. Now what do we need to do to make that work? Then, you know, obviously that's what everyone does now, pretty much.

Thorsten: The force underneath this that makes it possible is the model's just getting tremendously better in in the you know, last six months. Like I think it started with GPT 5.3 that started turn the tide and then 5.5. Fable obviously and now 5.6. I can send them off in an orb and say, What did I do yesterday? Oh, um, so we have this little sync button in the UI where you can click on and then it copies uh you can copy command, run it locally, and it syncs the changes locally.

Thorsten: If in any you want to try something by hand or modify something by hand and you want to get the changes out of the orb. I noticed in the web UI, if I open this, the tooltip over the copy button, it would already be visible even though I haven't hovered over the thing. Like a tiny, tiny thing.

Thorsten: I take a screenshot, and this is also for everybody listening, think how this relates to your tickets and your ticketing system and your boards. I took a screenshot, I opened a new thread. In ampcode.com like with an orb and I said, "this thing shows the tooltip with the screenshot whenever I open it, even though I don't hover" and I send it off.

Thorsten: That's the same amount of work I put into filing the same ticket. Except that if I had to file a ticket, I would maybe even write more in it so the other person doesn't have to grab the code. And then I think it was like I think I was on the call with the customer and then after the call I look and I see unread message and I go on it and it's fixed it.

Thorsten: Like I'm pretty sure it fixed it. And it says like here, I tried to confirm this. This is this.

Thorsten: Here in the storybook it's this. It was this attribute. And it did it.

Thorsten: And and it's just ... Yes, I could have maybe done it myself. That's what a lot of people say. Like, oh, you could have done this yourself.

Thorsten: It's just a one-line change. But this I just wrote the thing, sent it off, and it came back and did it. And that's a tiny, tiny example, but it can do this now with huge changes too.

Thorsten: Where you're you're like... For example, we we shipped um the Amp runners or `amp --no-tui` where you can run Amp the CLI in a headless mode without a TUI and it connects as a runner and then you can spawn new threads on this so you can you know, if orbs are the cattle then this is the pet case and you have your Raspberry Pi, which I have any construct threads on it. And I had there were some issues with it when naming conflicts and whatnot and an error message pops up and I send it off and then I said and then test it in the dev server and make sure it works. And it goes and it starts our dev server and it knows how to start multiple threads in it and it knows how to reproduce the error and then it takes its changes, builds a new version, and confirms that it actually fixed it.

Thorsten: And it's not like we have a secret sauce in our... It's not like we have like a skill, you know, that says, "here's how you confirm changes in an orb". The models are so good at this that like we we said like a year ago, one and a half years ago, we said like you're holding them back if you don't give them enough tokens.

Thorsten: And I think now you're holding them back if you if you keep them on your machine and ask them you know, like "only run in these four checkouts whenever my machine is online", you know? It's so that changed the game completely how good these models are. And we're still still incredibly surprised by it.

Quinn: So, you know, Devil's Advocate, some people might say, oh yeah, well, I've got my other agent hooked up in Slack. I've got Devin or Codex or Claude hooked up in Slack and it can make PRs. Like, are we saying that Amp's orbs are better or different or how do we see the future differently from everyone else?

Thorsten: I think a lot of other companies underestimate where this is going. I think they... Most people treat cloud agents as the more intelligent version of dependabot. You know, there's a pull request.

Thorsten: Then go look at the pull request and leave some comments and then do the following and then that's like it's it's deterministic harness for these models and event based and that's cool and you want this and you want automation and we're gonna ship this and it's already running locally and you know, Hitesh has some amazing stories to tell here. Um, but I think like others are more in this mindset of cloud agents, that's the thing you run on PRs. But I think in our case it's more... This is the obvious next generation of coding agents.

Thorsten: This is this is where this is going. But

Quinn: This is not... I've heard a lot of teams that say, "oh, that's for the nitpicks for the typo fixes or when there's something in the logs and it'll go and do the draft PR". But I think what's interesting and we have the luxury and Amp to do this, what if you just take it way further and you say where we where's this going?

Quinn: Let's let's jump ahead there and how can we make it so that actually, you don't need local dev anymore. And you don't want local dev anymore.

Thorsten: Yeah.

Quinn: And yeah, you know, just going deeper on that.

Thorsten: And go ahead, sorry.

Quinn: It's hard if you have not garbage collected a lot of the concepts in your product over the last year.

Quinn: They become obsolete really fast. I mean there's still some products that still have tab complete that are still based on editors that still have maybe half their users using that and just think of the distraction.

Quinn: So for us, our advantage is, thanks to everyone who uses Amp who's on board with us, we want to be the fastest to kill features that no longer make sense. We gotta earn your trust so that you actually follow us on the frontier. But if so, then I think that this new era when you can have parallel agents running I think if you're willing to be on the frontier here, if you're willing to make your code base work, your dev workflow, your team work with this, I think you're gonna get way more out of this leap than even, you know, going from like tab complete to agents.

Quinn: Um there's one thing though that is interesting about our own team and a few other teams out there that mean we have very different experiences. So on the Amp team, we all use Amp, and you can bet that our code base works incredibly well with all the features of Amp for like the dev preview servers via portals, which is something we'll ship soon and the setup script, it works really well. Like our Slack integration, which we're gonna ship soon, all these things. Um if you're using like 17 different agents at your company where, you know, each person kind of manages a different one's setup and, you know, it's gonna be janky.

Quinn: Like even the Slackbots are gonna be fighting with each other sometimes. Um also if you look at like the OpenAI team, they all use Codex. And Codex works so well for them.

Quinn: I imagine Anthropic, you know, using Claude stuff. It works so well for them. And it's this interesting thing too where up till now the coding agent has been a single-player tool.

Quinn: And that means that, you know, you can drop Claude Code or Amp or Codex CLI or anything, the desktop app in your repo is probably gonna do like a an amazing job. But you know, when it needs to actually run things in the cloud, there's a little bit more setup required. It's a little bit more collaborative.

Quinn: And so I wonder if this is going to mean that actually you're gonna make this decision as a team rather than as individuals.

Thorsten: Yeah, that's a good point. I do I think it's also gonna take some... The lift might be bigger to get people off the local dev. Um I think with you know

Quinn: Yeah, I think a lot of people listening to this probably want to get off local dev, but then they are probably thinking, how do I get the rest of my team off local dev?

Thorsten: Or "our dev env requires this and it would require this..." and "ugh I can't get it to work in something else and would require" blah blah blah and also.

Quinn: My company will never allow using orbs or remote VMs. And I think that there's not a lot of "nevers" left. I think a lot of them have already been broken in the last 12 months.

Thorsten: And to underline one point. I think most developers that I talk to underestimate by a factor of whatever, 10x, how useless the dev env, the local dev env now is. Like I did a poll internally here last week, I think, um where I asked how many of you like on the Amp team write code by hand?

Thorsten: And I put like three options, I think 99% then. You know, ninety to uh sorry, like how much is AI generated, right? Like 99% of my code plus or between 90% and 99% and then less than 90% or something.

Thorsten: And I think Camden hit the button you know say "below 90%". Like he's like, "I still write like 10% by hand". And then in I I pressed him and I was like, "what what type of code do you write by hand?"

Thorsten: And he he immediately he's like, "I actually now did you like it might be ninety plus like" and and then Tim and Lewis they were like "I haven't opened the editor in months" like "I don't do this". Lewis says like "I'm only in orbs now. This does everything I need".

Thorsten: Like, "why would I not like why would would I go back to local dev?" And I'm saying this because... Look, what was always the reason why you would never use Cloud IDEs? It's "oh it's on my editor, I cannot use my terminal, it doesn't have my dot files, it doesn't do this, it doesn't blah blah blah my shortcuts", like all of that stuff.

Thorsten: And using a Cloud IDE always felt like um what's it called? Like at the at the ferry, these little things where you drop a coin in and then you have to remote control the thing to get the the plushie toy out, you know?

Quinn: The grabber.

Thorsten: So like yeah, and then it's like this is what Cloud IDE felt like. But now it turns out I don't care about the text editor shortcuts anymore. I don't care about my shell tab completion anymore.

Thorsten: I don't care about the language server anymore. Like all of the local like what do I do? Like I send prompts to my agent and then review the work and I can I can do this in orbs.

Thorsten: Like they're really good at doing stuff in headless machines. Like they can use a browser. They can use playwright.

Thorsten: They can give you videos and screenshots. Like,,, They can we're gonna release this soon. Also, meta note, tiny meta note. We've been releasing stuff because we got sick of not being able to talk about what we use daily, and now I know this in this conversation alone we mentioned five things.

Quinn: I know... yeah

Thorsten: Then we need to get out because we use it internally. But so we're gonna ship stuff where they can uh run servers and you can watch them and review stuff. So I think a lot of if you start from first principles, a lot of people miss that a lot that the reasons why they think local dev is is required, they're obsolete.

Thorsten: Like a lot of stuff does not matter anymore. Like a lot of stuff. I get it for like some things.

Thorsten: If you do a lot of UI and you got a pixel perfect and whatnot and you want to play around. And I did this yesterday with some UI thing where it kind of wanna play around with some text and whatnot and see different versions of it. But for a lot of stuff.

Quinn: I mean you can still do that in an orb.

Thorsten: I know, I know. It was... In the end I didn't actually do it, but I thought I'm gonna go in and change the text and add a new line just to see the text layout.

Thorsten: But then I said to the agent, "I think we should change the text layout" and it came back with something that was pretty good. So I could have done it remotely. But obviously it requires some calibration, right?

Thorsten: Building up some instinct of what's possible and what's not possible. But my point is people underestimate or they overestimate how much they actually use the local dev env still. And they underestimate how good these agents are in orbs.

Quinn: Yeah.

Thorsten: And once you have them in orbs, they're async. You can start them and you don't have to watch them. You can come back to them any point you want.

Thorsten: You can trigger them from everywhere. The amount of paper cuts I fixed via phone by taking a screen. Somebody sends me a Twitter message like, hey, what is this?

Thorsten: I take a screenshot, drop it in a thing, and say that can you fix this and then push to main and it goes and fixes it. And that's just

Quinn: Actually so that's another thing. We push to main. And every time you say this or we say this, some people are like, "yeah!", and other people's like, "what? How do you review?" And I think that everything becomes so much slower if you're having to go to a pull request, get it reviewed by another human. And then if it's gonna be that slow anyway, then maybe you don't you know, if if that bottleneck is there anyway, what's the point of clearing up another bottleneck, just more stuff is gonna pile up in a review?

Thorsten: Yeah. Okay, you just... You just — what's uh what's the expression

Quinn: Triggered you?

Thorsten: yeah yeah well the the the ho the the bee's nest or the hornets... you just remind me of something... If you have orbs Sandboxes that are not your machine, and you have infinitely patient agents that run your test suite in and another the machine that's clean, you know, it's clean. You just checked out the whatever you only have your changes in there.

Thorsten: There's no other state lying around, and it runs your tests. Why should you push it to another branch and then wait for the tests to run again? Like the whole idea of CI is bonkers when you think about it, like you now have like these the sandboxes that are all the same that are standardized across your team. They have clean checkout and now you run the tests.

Thorsten: And then you're telling me you're pushing to a branch and then GitHub Actions runs again and then you merge to main and it runs again. What are we doing? Like you all CI is now part of the agentic workflow.

Thorsten: And we created CI because it was too slow to run locally, right? Like that's that's why we have it, like CI source, because we started with people running builds on their machines. Oh, they had different configurations.

Thorsten: Builds were not the same. Oh, they took too long. Let's put them to a build server so they can run for a long time.

Thorsten: Now you have reproducible builds, CI, continuous integration, blah, blah, blah, blah, blah. But now we have the change running somewhere else in these environments that are also clean and standardized. Why like the the value of CI is completely changed for me?

Thorsten: Like it's...

Quinn: Yeah. Agents need CI That's more needed than ever before, which is this kind of, you know, fast compute that, you know, runs these builds. But I mean there is an argument for well, you know, well you'd have to uh check all the pairwise potential conflicts and combinations and you know, you gotta test the actual integration of this, but you know you're not doing that by pushing to a feature branch and just running, you know, on that. And if you're running the merge and the head there, you know, whatever, like it can get complex.

Quinn: But there's this whole class of processes that just do not make sense anymore. And to your point about the ease of fixing bugs now, there's a whole class of bugs where you should be more willing to risk having a bug because it is so much easier to fix it. And that's a really uncomfortable idea for a lot of people, but for people that really, you know, look at this, it's not like the right target is a hundred percent uptime for everything.

Quinn: That's why, you know, you you have SLOs. I don't know. Everything is a trade-off and there's one side of the equation that has changed dramatically and the other side will adjust, you know, might take different amounts of time for different people or companies, but it will adjust.

Thorsten: And I mean, zooming out, you know, this is the podcast of predictions. I think the pull and we've seen this, like we've said this, the pull of agents is so strong. That it will change have a dent on technological choices.

Thorsten: And for example when the web came out as a platform, obviously the web had some downsides to native applications, you know, like it it's completely different thing, but it had huge advantages. It had the advantage of distribution. You could always update this thing you don't have to ship CDs you don't have to ship whatever downloads and whatnot like the packaging is completely different... Huge advantages from a business perspective, you know, the distribution channels and updating and updates and whatnot, less cost to to distributing it even.

Thorsten: It might have come at some technical disadvantages where, oh, you know, like uh on a website you cannot do multi-window stuff that you did in you know spreadsheet whatever excel like there's disadvantages and I think when we talk about stuff like CI is changing, pushing to main is changing, you wanna have agents in sandboxes, whatnot. I think there's... people might object with valid technological arguments and say, "yeah, but if I do this, then I can't do this". But I think it's the same as with the with the web that the pull and advantages of embracing this are so strong that I think you should start to think about how do you model your code base or how do you model your team or how do you even model your product to make to get the most advantage out of this, right?

Thorsten: Like if I were to create a new software company, would I pick a product or a business where I think I need five people bound to five physical machines and only those can work on it at the same time or will I pick something where I know that I can leverage agents as much as possible because they can constantly ship and do, you know, whatever. Like would I pick something with a six-month release cycle? I wouldn't. Like yeah.

Quinn: Yeah. Even the idea of how customers use your software - a settings screen and all of the you know complex abstraction layer that most code bases have to allow customization at the application level rather than through the code.

Thorsten: Yeah.

Quinn: On the margin... Now that your customers have an agent where they don't have to understand your whole codebase in order to customize your product if they had access to the code and if they could run it themselves. Well, then that means you probably need less configurable settings screens now because code is in a sense the ultimate settings screen.

Thorsten: Yeah.

Quinn: So that's what we are thinking of. Not only are we trying to take this idea of local dev is soon gonna be dead.

Quinn: And what does it look like when you have tons of parallel or but also what does that do to how you use and run software? It's not just how you it's changing how you build software, but how does it change how you use and run software? More on that later.

Thorsten: Yeah, we're not gonna reveal this, but I think orbs are just the first step in this, right? It looks still like orbs might still look like sandboxes to some. But I think I touched on this, but to be super philosophical about it.

Thorsten: Previously you had your code somewhere, you had your agent somewhere, you had your context, you had the change that the agent produced. Then you would have a branch that you would push the change to, your context that created a change would live somewhere else, and then you would copy it to a pull request maybe or you have it write the commit message. The computation that produced the change was a local machine, and it's three disparate places.

Thorsten: Like it did, it's not one thing. With orbs, I can now send you the link to my orb and you will have the computation that produced the change, the computation that still, and this is the preview part, that can run the change, like they can actually show you how it works. the the context that produced a change, everything that I put into it and all of the code.

Thorsten: Like it's a it's one URL and it's all in there. And it's not three different places. And if you combine the context in one, that allows you, for example, to take this to the next step and say like, oh yeah, I got everything here.

Thorsten: You know, it's like It's like somebody sends you a pull request and plus a cloud IDE that's ready to go and configured, and you just have to say, "now change this". Like it's a completely different game change uh way of working. So I think with all of the stuff that we're gonna release in the next few days and weeks, uh that will hopefully start to shine even more.

Thorsten: But it might take some people to like I think it's gonna take some people a while to see where this is going. Like some people immediately see it. Like we talk with some of them, like we talk with Evan on Monday or Tuesday.

Thorsten: The he like sees where it's going, but yeah.

Quinn: So we've talked to a lot of customers, a lot of people using Amp, and what were some of the things that people said, you know, that that didn't get it, where they're not quite there yet. What kinds of objections did they have to this?

Thorsten: I think one of the most common ones was um "we can't get our dev to work in orbs because it uses I don't know 15 services or whatever", or "we need to VPN into some things", or you know like their dev env sounds like a pet type of thing where it's really elaborate and whatnot and they don't see how it can work in orbs. I think that was one of the most common ways and uh common objections and um

Quinn: And so for people who are in that world walk through when they're gonna have a realization that they don't need to live under that limitation, why they're gonna have that, what they need to do to change. And if you're a dev in that position, how you can help your company see the light.

Thorsten: That's uh yeah, that's a tough task. I'll try. I'm pretty impatient.

Thorsten: So I'm like in all of these conversations, I have to really hold back so I don't go "Guys, this is dead. Let's just jump to the end thing". But I think what people truly have to see is how easy it is to spin up whatever complex environment you want in in even a headless thing.

Thorsten: Like it's you know, like I said this yesterday to a customer on the call. In orbs, in Amp, Amp's orbs, we have two scripts: .agents/setup, like .agents/setup and .agents/resume. Setup runs when the orb first comes to life.

Thorsten: And then it's snapshotted. Resumes every time the orb wakes up again, so you can keep stuff running. It's bash scripts, you can or whatever, you know, it's just an executable, you can do whatever you want in there. And I showed this to a customer yesterday with just a simple example of, you know, that was like a it wasn't that simple. Like it was a dev server running a website, and then I could see the live changes in the orb.

Thorsten: And I just asked Amp to update the setup script and it knocks out fifty lines of bash. And nobody that listens to this is surprised by this. But I think the big surprise is then if you say, now I want you to always spawn these two services and make sure that they do this.

Thorsten: And then I want you to be able to take screenshots and then I want you to walk me through this and then I want you to do this... Like all of that stuff where I would have previously thought this is stuff that I do locally by hand, they can do it on their own now. And

Quinn: yeah

Thorsten: I think people just need to see this. Like even with like a complex dev env I think if you tell the agent like get this to work in this thing, like I think it can like there's a lot of stuff it will be able to manage.

Quinn: And well by definition, you're saying a task that has uh an objective function because it's literally make this work in this thing. Yeah, make it so it's got a clear goal.

Thorsten: Yeah.

Quinn: And yeah, I had a really similar case just like two hours ago. I want to make it so that the dev server our dev server would running an orbs has better like seeded data initially. And it you know, just did that.

Quinn: And that's the kind of thing that would be such a pain if we were to do it also so that it works for local dev because. Oh, I don't know if Lewis has other custom data that he's used to working with locally or is it gonna clobber that? Is it gonna, you know, kick off on Hitesh's machine a lot of other compute? I don't know and once you don't need to worry about local devs, and think about it, if you're the central dev tools team who's responsible for doing like the local dev setup, you don't have to worry about all these different snowflake machines.

Quinn: You just have to target one environment and you get instant feedback and it's gonna work the same for you as it's gonna work for everyone else. If something doesn't work then you could literally see what didn't work or you can just tell the agent, find all of the other threads for my team where something in these setup failed, categorize those things, tell me which one is the most high impact to fix purse and fix it.

Quinn: And you can it'll just do that.

Thorsten: Yeah, that's a good point. It's and it's not just snowflake machines, it's Is this offensive? I don't know.

Thorsten: Like Snowflake developers? Like the you know, like everything. I don't mean anything bad with it, but I myself, you know, like I myself like like my at once upon a time, I think until like maybe two years ago, I was really big on like custom dev env and shortcuts and all of this and Vim configs.

Thorsten: But every dev tool teams dream is a standardized env. Everybody gets the same editor. Everybody gets the same machine.

Thorsten: I don't have to worry about the Vim users versus the Emacs users versus the VS Code users versus like they want a standardized thing. That's the dream, right? Less stuff to worry about.

Thorsten: But then you had to accommodate for people with different editors and whatnot. But now with orbs you can aim to have one standardized environment because the only way you interact with it is by giving instructions. It's not like you have to make five different editors work or different database machines.

Thorsten: Or somebody likes to run Postgres in Docker and the others like to run Postgres in postgres.app and the other guy likes a Postgres service installed via homebrew. You know, like all of that stuff. No, we can just standardize on this one thing and it's pretty good at it.

Thorsten: And then you obviously have the win-win uh you know thing going on of if it's standardized then the you can put it in a docs or agents.md file and the agent will knock it out of the park every time. Like it just knows. So that's I haven't even considered the standardization aspect.

Quinn: Yeah.

Thorsten: Because even if it's not standardized, the tokens will wash away the differences. Like the agent are pretty good at at making it work.

Quinn: Yeah. I I think it is probably harder if you're trying to say, well, internally we're gonna let people use Amp and Devin and Cursor Cloud Agents and Codex Cloud and Claude Cloud and all these other things, then yeah, you have kind of a mess and you're back in the old situation that you were in. Um but I think the difference is you don't need to be in that situation anymore. You can choose to just use one and I think it'll be really interesting to see if companies do have more consolidation than they had in the previous agent generation.

Thorsten: Yeah. Alright, shall we wrap it up to or any any oh you want to give a you should give a sneak preview teaser of um starts with "s", ends with "ubs" or "criptions". Yeah.

Quinn: Amp has sometimes been called the Porsche of coding agents. Of course now. We don't fit in the coding agent box.

Quinn: Everything is an agent now. And a lot of people said, I love Amp. It's so polished.

Quinn: It's you know so good, but it's so expensive. And yeah, if you're able to use the subscription like your ChatGPT subscription, that is an incredible deal. And we love OpenAI.

Quinn: I mean props to them. I think they have done so much to make agentic coding accessible to a lot of people. Um, now Amp will let you use your ChatGPT subscription with Amp so that whenever Amp is using GPT 5.6, which is now Amp's main model. Then it's just gonna draw tokens from your subscription. So that makes Amp way cheaper.

Quinn: And we're gonna come out with subscription plans for Amp itself. We've got the Megawatt plan at $20 per month, Gigawatt at $200 per month that lets you use basically as many orbs as you can. There's some limits, but the it's like 750 hours and a thousand hours and that lets you live the orb lifestyle like we are living and feel what it's like when you are not counting the minutes of compute where you can just launch these agents.

Quinn: And because the tokens can come from your ChatGPT subscription or your SpaceXAI subscription or, you know, we'll add more in the future, you don't need to be counting the tokens either. And you should go get those subscriptions, go get the Amp subscription, and all the tokens that are are coming from your subs are gonna be going way further, you're gonna be getting a lot more joy, a lot more out of them when you're using them through Amp. So we want to make these Amp subscriptions total no-brainers, launch them in beta by the time this is out and go try them out. And send us your feedback. This will make Amp accessible to a lot more people. So if you're one of those people that has wanted to use Amp or did in the past and is too expensive, then hey, welcome back.

Thorsten: Yes. And if you're reading if you're listening to this, just check your settings in ampcode.com.

Thorsten: Like I showed this yesterday to a customer, what's already possible, and they were like, oh my god, this is already this is so good. I'm gonna use you know they have multiple Amp accounts, so. Yeah, I think people will be excited.

Quinn: Um and we're back. The shipping velocity, uh like we've got so much. It's we're we're blocked on writing news posts that we feel really good about. And it's hard because it is so engaging to just build more stuff that we use and make Amp better for us and and for other people using it. But we gotta pull ourselves and write those news posts, get these things in your hands. We will do that, right?

Thorsten: Yeah, yeah. It's just to tease, I think, right now, close to the finish line, just to throw some words out. Automation, portal, slack, desktop, two things called desktop, mobile, um Puck.

Thorsten: It's called uh

Quinn: How do you pronounce it?

Thorsten: Puck? Do you pronounce it puck?

Thorsten: I I say p uh whatever. I guess we gotta...

Quinn: You say

Thorsten: no no I gotta You say puff. German.

Quinn: Puck. Yeah, it's a German. I'm an American. Puck.

Thorsten: Yeah, I need to figure this out. But so all of this is so close to the finish line.

Thorsten: And I think we now have a pile up like this week, where like everything is kind of piling up, but everybody needs like a little bit of uh a little bit of a twist, like a little bit of a last minute polish, and then we can all, you know, get it out.

Quinn: Yeah. And we hear it from a lot of people, they're like, there's a lot of stuff in Amp and we want it hear like how it all fits together and we know we want to be doing more of that. So we did like show the product.

Quinn: I mean we talked about it. The point of this is not to talk about the product, just talk about like what it's like to be in this crazy time of building. But we will show the product more and show yeah how it's actually changing how we work and what the workflow looks like.

Quinn: Because I think it's different from how a lot of other people are using their agents.

Thorsten: Yes.

Quinn: So Stay tuned for that.

Thorsten: Stay tuned. And also last last change we should make. Instead of calling it a Megawatt and Gigawatt, we should just call it Kilowatt and Gigawatt, but like the orders of magnitude are off, just so people write us angry letters like why can you this is $20 and it's $200 but the orders of magnitude are off? That will be good but all right.

Quinn: We can make that change.

Thorsten: Alright. Then we'll see you soon. So we're gonna record a bunch more of these in season two.

Thorsten: Um we don't know how many, but there's a lot to talk about, a lot to ship. So we'll see you soon.

Quinn: See ya.