Beyang
@beyang
Followers
10K
Following
16K
Media
763
Statuses
6K
CTO @sourcegraph, building stuff for devs: @ampcode
San Francisco, CA
Joined September 2008
There are many coding agents that use a model selector as their core architecture, but Amp is the only one with an agent-oriented architecture that treats agents as the atomic building block. The right core abstraction simplifies a lot of things in a good way.
This is why a model selector doesn't make sense. 1. Make a tool-calling coding agent 2. Make the tools better 3. Some tools become subagents under the hood 4. Different subagents need different models 5. Maybe even their own models 6. So what would a model selector even select?
1
0
10
If you've tried other agents and feel that they're overhyped, only good for small things, or you're still writing most code by hand + tab completion, I recommend trying out Amp. Join Build Crew, too. The community is curated and a really great forum for agentic coding tips.
2
2
21
Even if you’re using agents other than Amp, give the Amp review panel in VS Code a shot. I think it generalizes well to common usage patterns across coding agents:
ampcode.com
Review Panel
@beyang Just tried it out. I love that it works for commits that Amp didn't make too.
2
4
42
Read about it here: https://t.co/MwaYHggQGg It's live now. Update to the latest Amp VS Code extension and hit me up with feedback. Happy weekend coding!
ampcode.com
Review Panel
1
0
9
The bottleneck in coding is now the time it takes a human to review agent-generated code. But existing review UIs don't work well for agents. So we built a new one into Amp.
31
16
266
Quality ad partners allow us to support Amp Free without collecting data for training. Amp is now the only coding agent to support free inference without training data collection. That means you can use it for work—spread the word!
Amp Free: now with no training required, so you can use it at work. The Internet’s best business model (advertising) delivers once again: a $0 coding agent that meets the same stringent infosec standards of Amp’s paid `smart` mode.
5
2
59
People farming engagement will bounce forever to the next shiny tool. People who just wanna build can just use Amp. We use models from all the major model labs, but intentionally in specialized subagents (the oracle, the finder, the librarian).
theres a million views making content about how you unsubscribed to {ant, oai, gemini} and subscribed to {ant, oai, gemini} cause the new model is better
1
1
29
UX matters more than ever. The model is not the product—the product is the product.
A friend just texted me: "Amp is so good and I love that I don’t know why." 😬😅 People cite Amp's multi-models (eg Sonnet + GPT-5 oracle + fast/free models), polished CLI & editor extension, thread sharing, etc. But...there's also an intangible trust that people seem to have
2
0
45
I hope more teachers and professors will come to view AI as a tool rather than an adversary. We've had a few professors reach out about getting Amp into their classrooms and are currently figuring out how to teach Agentic Coding 101. This is the way!
0
0
10
It became something of a meme where we supported compaction but would constantly warn folks from using it. Handoff is a better, more intentional way to extend context for long running features in Amp
Today I ripped out /compact from Amp. I think we found something better: /handoff Here is me and @sqs talking about why
2
2
28
We think Amp users are best served if we focus on building, not picking fights on Twitter. But a competitor that passes inference costs to users is spreading falsehoods about Amp Free. Help us spread the word: Amp Free covers cost of inference with ads. More to share soon!
@inferencetoken It's not dependent on the model provider (but Cline free inference is?) Amp Free (try here btw: https://t.co/qHUrbnDq3o) is dependent on the self-interest of devtool cos wanting to reach devs. Please be truthful when you tweet about us—we have never spread falsehoods about you.
4
7
52
You don't need to write an MCP server to give custom tools to Amp. 1. Tools are defined in vanilla JS 2. Have Amp write them for you, as Dario does here Custom tools can tell Amp how to run tests in a specific way or use specific utilities you have in your dev env.
We've made Amp toolboxes better: • AMP_TOOLBOX supports multiple toolboxes, like PATH • `amp tools make` creates a new, ready-to-edit tool • `amp tools show` inspects it • `amp tools use` tests it It's now super easy to create a new tool and get immediate feedback.
3
1
22
The thing I appreciate about Jason’s courses is they are NOT pay-to-teach. Students pay for the courses (he does give some away for free), and he just reaches out to the people who are authoritative on a given topic. Wish this were the rule and not the exception!
You guys are getting paid?
0
1
13
Anyone else find it both amusing and concerning that spreading the word about dev tools now involves paying someone in the social media circus $$$ to shill or joining the circus yourself? Does this incentivize the spread of good technology? Maybe there should be a better way...
9
2
48
This was one of the most often requested features in the Amp TUI. You can now edit earlier messages and also restore the code to the state it was in when that message was sent. Didn't like where a thread ended up? Reset and resend from an earlier checkpoint:
We added editing and restoring to previous prompts in the Amp TUI. Tab and shift+tab to jump between messages. E to edit, R to restore. This is one I really missed from the Amp VSCode extension.
2
2
22
The Amp finder, the subagent that searches through code to surface relevant context, is now 50% faster with no effective loss in quality.
7
10
102
There's a new subagent in Amp called the Librarian. Its job is to locate relevant context in all the libraries you might be using in open source or your private dependencies. Amp's accessible context now extends to the entire universe of code: https://t.co/6zgXMo8EZ2
8
13
145