The alpha, twice a week
That was 3 min.
Get the best new moments in your inbox every Wednesday and Saturday. It's free.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
AI & I · 3 min · Oct 29, 2025
The next coding models may run autonomously for days
Boris Cherny says their longest-running agent tests reach about 30 hours and predicts the next model will run for days. Longer runs will require remote execution and ways to monitor parallel agents.
Creator of Claude Code, Anthropic
Created Claude Code at Anthropic and leads its team; previously spent about five years at Meta.

Dan Shipper, Host · Co-founder & CEO, Every
Sold his first startup, Firefly, to Pegasystems; writes Every's Chain of Thought.
About this segment
Boris Cherny says their longest-running agent tests reach about 30 hours and predicts the next model will run for days. Longer runs will require remote execution and ways to monitor parallel agents.
From “The Secrets of Claude Code From the Engineers Who Built It”, Oct 29, 2025. Watch the full episode
Answers “What will AI look like in the next few years?” See all 53 answers
Thanks. We'll look at and fix what's off.
“And now we're on the order of double-digit hours. I think it's like the last model is like 30 hours, something like this.”
Boris Chernyat 46:23 · AI & I46:23Takeaways
- Their latest long-running tests reached roughly 30 hours.
- Boris predicts the next model will run for days.
- Longer runs raise questions about containers and monitoring.
In this episode
Full episode- 26:11Turn past code and mistakes into compounding leverage6:56
- 21:43Use parallel agents to migrate code and filter false positives4:34
- 2:20Claude Code began as a lightweight terminal prototype4:44
- 15:46A throwaway build can clarify a feature before planning4:43
- 33:09Latent demand shows what to build next3:31
- 40:57Agentic search avoids stale indexes and simplifies deployment4:10
45:01The next coding models may run autonomously for days2:46
- 7:57Anthropic's internal dogfooding gives Claude Code rapid feedback6:08Up next
Transcript
From the transcript AI & I publishes with the episode.
- 45:06
So you guys started this sort of CLI, this CLI paradigm shift. Do you think the CLI is the final form factor? Are we going to using cloud code in the CLI primarily in a year or in three years? Or is there something else that's better? I mean, it's not the final form factor, but we are very focused on making sure the CLI is like the most intelligent that we can make it. And that's as customizable as possible.
- 45:32
You can talk about the next form factors. Yeah, I mean, Kat's asking me to talk about it because no one knows. This stuff's like, it's just moving so fast, right? No one knows what these form factors are. Right now, I think our team is in experimentation mode. So we have CLI, then we came out with the IDE extension. Now we have a new IDE extension that's like a GUI. It's a little more accessible. We have addclawed and GitHub, so you can just addclawed it anywhere.
- 46:00
Now there's Cloud on web and on mobile, so you can use it on any of these places. And we're just in experimentation mode, so we're trying to figure out what's next. I think if we kind of zoom out and see where this stuff is headed, I think one of the big trends is longer periods of autonomy. And so with every model, we kind of time how long can the model just keep going and do tasks autonomously and just in dangerous mode in a container, keep auto-compacting until the task is done.
- 46:27
And now we're on the order of double-digit hours. I think it's like the last model is like 30 hours, something like this. The next model is going to be days. And as you think about kind of parallelizing models, there's kind of a bunch of problems that come out of this. So one is, what is the container this thing runs in? Because you don't want to have to close your laptop. I have that right now because I'm doing a lot of DISP.
- 46:49
I don't know. I've only heard, I've only read it, but DSP or DISP prompt optimization and like it's on my laptop and it's like I don't want to close. I'm like in the way though, like with my laptop open because I'm like I don't want to close it. Yeah. Yeah, that's right. Yeah, we like visited companies before, like customers. Everyone's just like walking around with their like quad codes. Is this running?
- 47:10
So I think like one is kind of getting away from this mode. And then I also think pretty soon we're going to be in this mode of like clods, monitoring clods. And I don't know what the right form factor for this is, because as a human, you need to be able to inspect this and see what's going on. But also it needs to be quad optimized, where you're optimizing for bandwidth between the quad to quad communication.
- 47:31
So my prediction is terminal is not the final form factor. My prediction is there's going to be a few more form factors in the coming months, maybe like a year or something like that. And it's going to keep changing very quickly.