Building @ampcode, founder @sourcegraph
Joined September 2008
- Tweets 6,811
- Following 1,949
- Followers 12,051
- Likes 18,738
844 Photos and videos
❤️
Jun 12
Surprising take: @AmpCode Web UI is the first AI interface that’s actually pulling me away from the CLI.
Not because it tries to do everything. Because it gets out of the way. Flexible, controlled, no useless bloat.
It’s not complete yet, but the direction is right.
My only asks:
• start threads directly inside server/local workspaces
• choose models per thread
• create isolated sandboxes per project
• keep keyboard ergonomics at CLI level or better
Right now I start a thread from the server, send “hi” just to make it appear in the Web UI.
When that flow lands, this thing is going to be hard to beat.
Knowing what the AmpCode team has already shipped, maybe this take shouldn’t be surprising at all.
10
2,929
Intelligent Distributed Executors
29
2,999
Fable in Amp instructions:
Jun 9
amp update
amp plugins add --auto-update @amp/fable-mode
amp --mode claude-fable-5
1
10
1,477
Evolution of my primary coding agent UI:
January: fully in the editor
February: VS Code terminal diff view
March: Ghostty
April: back to VS Code terminal (more precision code)
May: Amp GUI (80% PWA, 10% browser, 10% mobile)
June: who the heck knows, everything is changing!
11
1
88
5,781
A new UI for driving Amp from your desktop or, as I often do now, from your phone:
Jun 4
We rebuilt Amp's UI so you can watch and drive all your Amp agents on web, mobile, and CLI.
ampcode.com/news/agents-ever…
2
1
26
2,392
Beyang retweeted
Jun 1
I love when technological progress kills a feature.
The models are so good that (A) this feature is too risky and (B) it's not needed anymore to convince people that agents are good.
ampcode.com/news/end-of-publ…
17
10
153
15,153
There are plenty of coding tasks where you need to be mindful of your mental model and actively steering the agent to implement a clean architecture. For those, you want something fast and interactive but still smart and competent.
What we learned from the previous rush mode is that cheap tokens do not necessarily mean cheap in terms of total cost or time spent. If you haven’t used rush in awhile, try it again!
May 21
We made rush actually rush.
It is now GPT 5.5 with no reasoning tuned for smaller, bounded coding tasks where higher reasoning is overkill.
It’s almost twice as fast and outperforms the previous Haiku-based rush on our internal evals.
4
4
66
10,075
If you're a builder who's tired of the AI discourse being about job loss, class warfare, and doomerism, come work with us as we figure out how to deploy AI to amplify the agency of humans working in major industries.
Amp Labs: small teams of the best software builders, working inside the most frontier-oriented company per industry and region.
ampcode.com/news/amp-labs
11
2
90
7,563
As coding agent usage caps continue to tighten, Amp's pricing remains the same.
The "Porsche of coding agents" is increasingly near price parity with alternatives. Rather than switch tools every 3 months, Amp users can use one opinionated harness that's built for the frontier.
7
2
82
7,502
Wait til you hear how this was built
May 7
Markdown rendering in Amp is now 2x faster end-to-end, thanks to a new parser we (ahem...Amp) wrote today. This parser is 100x faster in most cases, and in large files 200x faster. Uses struct-of-arrays and buffers for low allocations and lazy string materialization.
3
39
9,143
A completely rebuilt-from-the-ground Amp TUI is out today!
3
1
60
3,697
New model for Amp’s smart mode
Apr 25
Amp's smart mode now uses Opus 4.7.
It's very, very good.
And we could remove three tools.
ampcode.com/news/opus-4.7
1
15
2,279
Note on GPT 5.5's data retention requirements, which appear to apply broadly to agents using this model. If you're willing to waive ZDR (zero data retention) by OpenAI, you can use it in Amp with these instructions.
(Still no training on this data, even w/o ZDR, in Amp)
Apr 24
You can use OpenAI GPT-5.5 in Amp's deep mode with this in ~/.config/amp/settings.json:
{
"amp.internal.model": { "deep": "openai:gpt-5.5" }
}
Will be the default soon. Why not yet? Because...
For now, usage of GPT-5.5 uses OpenAI's Safety Retention Policy, which means non-zero data retention by OpenAI in the (~0.05%) case of inputs that OpenAI's classifier identifies as severe cybersecurity abuse risk. We're working with OpenAI to lift this requirement, which as far as we know applies to use of GPT-5.5 in other agents as well. (Has anyone seen this disclosure from other agents?)
1
13
3,366
Wow this looks slick!
Mar 27
We built the fastest browser automation stack for AI agents.. 464 KB, 3ms cold start, 16% cheaper per agent loop than anything else out there.
Here's how it started and why it exists.
Shoutout to @AmpCode for the credits for the initial start of this project!
1
7
3,212