Quinn Slack

@aaronluannguyen @jeroendee 🫡😍

@sqs @jeroendee While I’m still struggling with costs on my personal projects, I’ve been gleefully flying with Amp at work and it’s night and day with the other available agents. You all have hurt my willingness to use other agents for my personal projects tbh 😅

@sqs Quinn, wouldn't AMP benefit heavily from an LLM that you trained yourself? Lately the feeling is creeping up, that the dependency on the frontier models is not sustainable. The subsidised bubble is going to burst. Would this be financially attainable?

@jeroendee (I define "best agent" as the one chosen by the top 1% of devs, the ones that push the world forward, AND the one that has a great and profitable business model.)

Being in the middle, with an okay agent and an okay model, is fatal. But there's a lot of opportunity in being on the extremes. Amp has always proudly been on the other extreme, building the very best agent that can use all/any/best models with no constraints, picking the winners at the model layer. That is our 100% focus. If we had USD$100B+, we could train a model and it would be really really good. But we don't, so we won't. If we did, too, then we probably wouldn't build as good of an agent. (I know there are some reasons why tight coupling of model<->agent can improve the agent, but I don't think those dominate the outcome.)

@xlatentspace @AmpCode @thorstenball Awesome!

@sqs @AmpCode @thorstenball Fixed! Thank you kindly!

@AmpCode @thorstenball @sqs I've run into an MCP related bug where amp is stripping fields even if additionalProperties: true. I am building a fairly large surface area local MCP server that has custom input variables across a large number of scenarios. It's critical to not publish the full set of fields in the top level tool. This design is confirmed to work in Claude Desktop and claude code.

We want to change the workflow more often to match what we think the current best way to use agents is. We need to do that more. We've been growing like crazy and dealing with that, plus doing some big improvements to the overall system, but that'll ship soon and then we'll be back to this. And then if you are using Amp in a way we don't recommend, you won't get the free $, and if you are using it in a way we do recommend and helping us build the future through your feedback, we'll give you more free $ to offset the growing # of tokens you're probably using. TBD on how we'll do this, and to be clear, that does mean that many people will get less, I don't want to sugarcoat it.

@sqs @AmpCode Thanks @sqs appreciate the reply, I have read that when you guys published it earlier, when you say "free daily grant to be more available and more generous for people using Amp in the recommended ways" what does the recommended way mean?

I truly do love @AmpCode and what they offer and free $10 but lately it's just been impossible, my entire $10 free credit goes to thinking and looking and thinking and look and thinking and looking and thinking and looking and thinking and looking it's insane...

@erd0xbc 😎

@sqs Neo UI! Time to see what's cookin' chief

I love the feeling of having 12+ hours of focus work ahead of me with zero interruptions

Bike ride to/from dinner

@singkid9527 @AmpCode That top screenshot is not from Amp. The bottom is. What is the top screenshot from? And why did you think it was from Amp?

Hey @AmpCode @sqs why does it say "credit limit reached" when my dashboard clearly shows $10 remaining?

Useful technique in Amp with GPT-5.5 (deep mode) to make small independent commits from a big uncommitted changeset: "git stash the uncommitted (including untracked) changes, then reapply this refactor that you performed on HEAD, then git stash pop and fix merge conflicts"

@JamesNguyen868 It’s in for many users, rolling it out more, and you can get it now at: x.com/sqs/status/204…

@sqs When gpt 5.5 for amp Quinn?

@tinvaan nope just at home

@sqs flying to Singapore again?

@dropalltables put one in Marin pls

Fundraising for Dataspa now!

@jonathangrahl @AmpCode Probably early next week, maybe tomorrow, will email you soon

@jonathangrahl @AmpCode Let's get you in the automatic compaction beta ASAP. Put you on the list. This will mean never needing to think about context windows or handoff again (as long as you keep 1 thread to 1 logical task).

I feel like my @AmpCode deep^2 usage has to change significantly for me to stay productive. I keep running out of context even with a single question in our larger monorepo. Time to try Smart mode again?

@TheEthanDing This is good.

@j_schwartzz Yes, many individual users and SMBs use subs, but someone’s gotta pay for all the fixed costs. That’s enterprises. Generally, the labs make it so enterprises can’t use subs. It’s very logical price discrimination.

@sqs The vast amount of Claude or Codex usage is not on subs? Is that true? > token pricing is viable and a sign of high demand for a product Agree!

So far I'd say that both are doing quite well, but neither are exceptionally better than default Codex. The task I gave: - Implement a plan to build a server according to a spec that was already written - Migrate a sibling repository (written in TS) into the current repo, translate it to Go This isn't a super easy task. I didn't give it any new skills or guidance whatsoever. I wanted to see what the out-of-the-box behavior felt like. AmpCode jumped right into it, started scaffolding quickly. It left the code incomplete, and has been stuck "running tools" for a solid 20-30 minutes now. I'm just letting it ride out. I bought $5 worth of credits, and 132K tokens cost me $3.28. One more equal sized session, and I'm on pace to spend >$200 a month in Amp credits. I could see myself using AmpCode, but pay per token (when other model providers offer very generous token subsidies) ain't it for me. Factory created a plan first, which I liked, and I enjoyed how it started implementing the plan, and showed me the exact step by step to-do list and when each step was finished. Gud UI. I can't tell how much of my $20 monthly plan has been used in this turn. Similar to Amp, it's stuck on the last "summary to user" to-do item, after it already finished the summary. With Factory/Droid, I feel way more likely to use because it seems plausible (and somewhat allowed by OpenAI) to use my Codex oauth subscription as a backend for Droid. That's a solid option. Lastly, I had codex compare the two implementations of both to see which is better, and it chose Factory/Droid. Safe to say, both are fantastic tools, but neither have been wildly better than Codex to the point where I can't see myself going back. At least for the projects I'm working on, all coding agents need some love and guidance, token subsidies make a bigger difference to me now than better harness