Thorsten Ball (@thorstenball) header
Thorsten Ball (@thorstenball) profile picture

Working on @ampcode. Writing https://xitter.gq/t.co/0SZD6dbLes. Once authored https://xitter.gq/t.co/arW0hnVET0 and https://xitter.gq/t.co/RN9xXOzhON.

Großwallstdt, Bavaria, Germany
Joined November 2011
Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)Photo by Thorsten Ball (@thorstenball)
Most predictions I see are still way too conservative. Here's mine
556
949
415
9,485
2,558,174
Crazy how the well these models now handle multiple questions and requests being sent at the same time. Reminder: forget most of the context management practices you thought were important 4 months ago.
47
8
7
229
13,422
Really miss the $1.50 pizza in NYC. Sitting on the curb, eating a slice, seeing a rat, that's the dream.
18
7
9
84
6,855
Dude, this is a lot of fun. Passing the Apple TV remote back and forth is a lot of fun for a turn-based game. Surprisingly easy to deploy to: select TV as target in Xcode, hit Cmd-R.
Building a Worms Armageddon clone for my daughters to play on tvOS. Our Mac sidecars in Amp are amazing. (Gemini 4 Neon!)
18
4
2
117
9,780
Hey, this blew up more than expected. Very close to 1 million subscribers now. Roughly 990k to go.
Programming alone isn't enough anymore. More than ever, it's time to put more than code into software.
11
190
12,066
Building a Worms Armageddon clone for my daughters to play on tvOS. Our Mac sidecars in Amp are amazing. (Gemini 4 Neon!)
23
1
78
14,116
Thorsten Ball retweeted
Me and a bunch of your favorite programmer-writers (Thomas Ptacek, Thorsten Ball, Hillel Wayne, etc.) are judging this neat writing contest. I'm looking forward to your submissions. Let's help the kids. RT for reach. advicewanted.dev/
6
36
2
258
11,904
Thorsten Ball retweeted
It's dangerous to look at and type into your agent while driving. It's also not exactly joy-inducing. Now you can do it all with voice and the occasional big picture or button. It's way safer and actually quite productive and joyful. ampcode.com/news/puck-in-the…
13
6
7
105
68,991
Programming alone isn't enough anymore. More than ever, it's time to put more than code into software.
29
20
14
298
36,903
Thorsten Ball retweeted
It still feels to me as if dropping through an event horizon. That is, there is nothing specific about any given moment, but the rate of change and weirdness is constantly speeding up, and we won't ever return to the before times.
87
132
15
1,830
36,347
I wish I could connect my agent to Spotify. Here's 10min of me rambling on about music and what I want to feel when I listen to it in the car, now please make me a playlist.
45
142
13,817
"Claude flew, and danced, and sparkled near that finish line."
JUST IN: Riemann Hypothesis has been solved. alpo.ge/remark.pdf
6
17
8
299
21,492
"But you won't be able to program without an agent." "How well can you program without the internet? And how often do you?"
58
29
16
441
19,203
This is something that comes up again and again. If, say, batteries get 200% cheaper, it's not just that the same products you had before become cheaper; you will also get entirely new products that weren't feasible before. Same for code.
Replying to @KaiMeyer
Sure. But I think you're still underestimating speed. It's like saying "oh, you made computers smaller? doesn't matter to me, because I have enough space on/under my desk" But small computers result in computers being in lightbulbs, in phones, in earbuds, in glasses, in dish washers, ... Some changes in a single attribute create new categories of use. We can write a lot of code faster now and that not only means that we now create the same things faster than before. It also means that things are now feasible that weren't before: additional tooling to prove just a single change correct, prototypes no one ever would've built, papercuts that would've taken too long, delightful moments no one would've spent time on, ... Not to speak of all the tooling where *writing code fast* now even makes the product possible: you can literally have an interactive dashboard where everything you talk into a microphone gets turned into SQL queries by an agent; code mode (!); search (!).
21
6
2
165
9,132
Replies are full of people super happy with spaces. Just goes to show that—
Spaces might be the most controversial Amp feature we've ever launched. A lot of you immediately said: this is great, exactly what I wanted, love you, it's orbin' time my brother, etc. But others asked: "How do I use spaces? Is this not for me? Is it only for multiplayer?" What I find interesting about all of this is the meta observation that the conversation with a single agent is, in some way, "solved." Not a single question about that. Instead everybody is asking about how you hold the transcript: when do you start a new thread, do you use subthreads, when do you dive into subthreads vs. staying at high level, how do people collaborate with it, how do you organize them? And spaces hit right into this spaghetti web of opinions, approaches, and thoughts. And part of the confusion is ours. We thought of spaces like Slack channels: we have a Design space, a bug tracker space, and I have one for Boxes, where I shipped Boxes and now maintain and monitor them. And just like in Slack, the work still happens in threads, not in the main chat. But a lot of people interpreted spaces as chiefs of staff (and, frankly, our announcement didn't help): you talk to the main agent and never dive into the subthreads. That leads to disappointment when the main agent doesn't reply and they have to go into the subthread anyway. UI suggested this, agent does that. That, of course, begs the question: "well, Thorsten, if you don't have answers to all of these questions, why'd you ship spaces?" Three years ago I would have said we shouldn't have shipped it and polished and iterated more. But in 2026 everything can change in three weeks with a new model drop. It's, I think, important to hit the wave while you can and collect feedback. (Of course that was always true but with AI even more.) But it's also frothy and caused confusion (and delight!) and, in hindsight, I wish we'd underline the exploratory nature a bit more. We all believe the workflow of the future is a high abstraction layer, not micromanaging threads. But there's no denying you still have to go in, review things, and make sure what you told the agent to do is correct. Now our #1 goal internally is to match the UI and the agent's behavior to that.
8
2
31
5,482
Spaces might be the most controversial Amp feature we've ever launched. A lot of you immediately said: this is great, exactly what I wanted, love you, it's orbin' time my brother, etc. But others asked: "How do I use spaces? Is this not for me? Is it only for multiplayer?" What I find interesting about all of this is the meta observation that the conversation with a single agent is, in some way, "solved." Not a single question about that. Instead everybody is asking about how you hold the transcript: when do you start a new thread, do you use subthreads, when do you dive into subthreads vs. staying at high level, how do people collaborate with it, how do you organize them? And spaces hit right into this spaghetti web of opinions, approaches, and thoughts. And part of the confusion is ours. We thought of spaces like Slack channels: we have a Design space, a bug tracker space, and I have one for Boxes, where I shipped Boxes and now maintain and monitor them. And just like in Slack, the work still happens in threads, not in the main chat. But a lot of people interpreted spaces as chiefs of staff (and, frankly, our announcement didn't help): you talk to the main agent and never dive into the subthreads. That leads to disappointment when the main agent doesn't reply and they have to go into the subthread anyway. UI suggested this, agent does that. That, of course, begs the question: "well, Thorsten, if you don't have answers to all of these questions, why'd you ship spaces?" Three years ago I would have said we shouldn't have shipped it and polished and iterated more. But in 2026 everything can change in three weeks with a new model drop. It's, I think, important to hit the wave while you can and collect feedback. (Of course that was always true but with AI even more.) But it's also frothy and caused confusion (and delight!) and, in hindsight, I wish we'd underline the exploratory nature a bit more. We all believe the workflow of the future is a high abstraction layer, not micromanaging threads. But there's no denying you still have to go in, review things, and make sure what you told the agent to do is correct. Now our #1 goal internally is to match the UI and the agent's behavior to that.
59
4
3
200
16,853
There you go. Amp stack, the thinnest of them all. Just tell the model what you want.
I removed 90% of my skills when I moved to amp code. I have not needed to revert that choice in the couple months since the move.
7
3
58
5,611
Thoughts after talking to lots of engineers in London in last couple of days: • Everyone's moving into the same direction and getting more and more agent-pilled, but all on individual speed. No group of engineers right now is uniformly agent-pilled. • Becoming agent-pilled is a one-way door. You only realize this once you've gone through. Many think they "get it" but haven't made the switch yet. • Wild how many orgs are still treating agents as "thing that writes code on my computer" instead of robot-engineer that can help with whole SLDC. • Teams are drowning in PRs and code review. No process has been changed yet, even though this is the biggest dev productivity lever of last 30 years, engs still work off tickets like nothing happened. Wild. • Barely anyone realized yet that, hey, even if the model isn't good enough to write the code the way you'd write it [doubt it], they can surely write all of the peripheral tooling to speed you up. • Very few engineers have imagination to come up with better tool. Instead of reviewing code in diffs, line by line, think: if I could wave a magic wand and make tooling disappear, what tooling would I want to gain confidence in this change being correct? Full instrumented/traced end-to-end runthrough? Demo videos? In-production probes? High-fidelity and anonymized production data fed through test runs? New testing environments with the greatest debugging MCP support one can imagine so agent can inspect everything? It's time to dream big. But seems like most people's dreams don't reach further than "this agent works off the tickets assigned to me 10x faster than I can."
116
88
10
1,607
101,775
Thorsten Ball retweeted
If you are in SF tonight come hang out with the @AmpCode and @laravel teams! I’ll be there and you can see my glorious fish face in person. Also @sqs and @beyang will be there, too. Which, is fine, I guess. Well show how teams build together with Amp with practical tips. Plus, a live build using Laravel Cloud. I can’t make any promises about the perpendicular Golden Gate Bridges making a cameo. Sign up here: luma.com/laravel-1x0r
4
4
1
45
2,962
New episode is out!
New episode: Raising an Agent S2E8, “Features Are Meaningless Now” @sqs and @thorstenball return to discuss shipping before it's polished, why a feature list isn’t a moat, agents modifying the production database, and whether to kill the settings screen.
3
2
55
5,290
We talked about this at the Amp meetup yesterday and everyone agreed: Elaborate pyramids of skills that are invoked deterministically by hand, by a human, seems like an anti-pattern.
We don't use anything instead of poteto mode. And it works fine without :) I also personally think that invoking skills with slash customs is probably an anti-pattern: - you want the skill to be discoverable enough without a slash command - if it's a highly specific thing that's deterministically invoked, it makes me wonder why one uses the model as a workflow executor vs giving it a high-level goal and figuring it out - codebase should be self-explanatory and legible to the agent without deterministic invocation by human. If human has to tell agent when and how to review with highly specific skills, something probably went wrong somewhere else before
28
2
5
259
19,273
Don't get me wrong: we have skills, I use them, I like them. But there's a world of difference between having a using-dev-server or writing-commit-messages skill in the codebase vs. use this skill to create plan, use that skill to validate plan, use this skill to split plan up, use this skill to distribute to subagents, use this one to sync to tickets, now review with this one and clean up with that one. If you're doing the latter I suggest trying to go a week without them and just telling the model What you want and encoding more and more helpers and notes in codebase.
4
1
25
2,623