Ian Lyons

@just_be_dev @tommoor @linear Tremendous. You guys did a great job on the agent, fwiw

@ianlyyons @tommoor @linear Nope, should be generally available. I tested it out on a few repos earlier.

@linear I'd really, really love it if I could chat with the linear agent in a github comment, similar to the claude github integration. Just being able to CC it to make a follow up issue would be huge.

@tommoor @just_be_dev @linear Do I need to do anything to enable it? I was thinking about this literally this morning

@just_be_dev @linear No need, I just enabled for everyone – happy Friday

@thsottiaux If you start drafting a comment in the diff viewer, then start drafting another comment prior to submitting the first comment, in some scenarios the text from the first comment shows up in the second comment, which is quite confusing.

It’s the little things that matter, what are some small papercuts you have noticed in Codex? We’ll fix as many as possible in the next week.

@namespacelabs @NEA @SusaVentures congrats!! we've been big namespace fans for a while... the support and responsiveness is unparalleled, and i continue to refer people who are looking for better solutions (and increasingly devbox/agent sandbox solutions!)

Today, we're proud to announce $23M in Seed and Series A funding led by @NEA, with @SusaVentures and burst.vc. Code workloads are unlike anything else: bursty but latency-sensitive, incremental but highly parallel, demanding reproducibility, provenance, and security all at once. No existing infrastructure was built for this. So we built Namespace. A compute platform designed for code from the ground up, across software and hardware. namespace.so/blog/series-a

@BrennerSpear Idk if you’ve tried @AmpCode, but this is their bet, and it’s good. The tradeoff is: when you blend the best model for every purpose, the result is definitely better. But you lose the massive subsidies that all the token providers offer for their own tools, so it’s much more $

In the short, medium, and long term Will having an orchestrator across multiple model providers (Claude, Codex, Gemini) be meaningfully more effective that an orchestrator that limits you to only their models (like Claude Code)?

@ctatedev this is one of the libraries that i am upgrading with each new minor release. the agents genuinely seem very good at driving it; kudos to the team

Diffing now available in agent-browser Compare pages with snapshot diffs or pixel-level visual regression Verify actions with up to 90% fewer tokens, catch visual regressions, monitor page changes, and compare across environments

okay this is getting out of control

@mitchellh @cablelounger ...please? the hero we need but don't deserve

@cablelounger No lol

Mitchell is going to launch a GitHub competitor 👀

THE UNBRIDLED GREATNESS OF TRAINS WILL ENDURE FOR A MILLION YEARS

@sqs And it feels… faster? More correct right away? But the window is very small, so my threads are constrained to more focused issues. Maybe a good thing, though…

Interesting as in trending even cheaper than Gemini per thread

Latest Opus cost numbers (per thread not per token) are very very interesting after almost 36h of widespread use in Amp. Will share soon…

@thorstenball Some nuance and blah blah but… overall my impression is better/smoother than Gemini, with none of the annoying reasoning-in-comments behavior

You can now try Opus 4.5 with us in Amp. We've been very impressed with it. This is big.

@sqs @AmpCode Yeah, totally agree -- especially the point on different roles! How do you envision this working? (If we were to, say, build this tooling in advance of Clicky being available...)

This is a key thing we need to get people to set up. I think you need to be able to give the agent pre-authed access to a variety of user account types (eg admin, individual user, team member) and that’s a step the user needs to verify. If the agent needs to figure it out each time, or the user needs to give a session cookie, it won’t be reliable or autonomous enough.

Meet Clicky, the browser use subagent in @AmpCode, which records live screencasts with voiceovers of its changes, to prove the correctness of its work to you. Clicky gives Amp another powerful feedback loop to write code better and more autonomously. WIP, will ship when ready.

@realhasanshoaib @cnakazawa Some of the LSP stuff still isn't there: import sorting/suggestions, semantic highlighting, blah blah. But overall quite good!

@ianlyyons @cnakazawa I was considering moving. What would you say is still missing from tsgo?

Once tsgo supports import suggestions I’m going to switch every project over.

@TheRohanVarma I'm interested!

We’re beta testing an experimental UI for managing coding agents in Cursor. Who wants early access 😬

@GeoffreyHuntley @namespacelabs +1000 to this. We’ve been working with namespace for a while now and the product is great and stable, and the team is shockingly responsive to our weird requests.

folks, if you want a replacement for GitHub Actions which still uses GitHub actions just switch to @namespacelabs - lightning-fast builds with great caching - best in class observability - remote terminal into jobs - oh, it's cheaper than MSFT by miles

@sqs @OpenAI Have you heard that it’s generally slower than other models? Anecdotally it seems much slower on most responses — not sure if it’s infra scaling issues or just a model trait

GPT-5 is also similarly faster in Amp now. Thank you to @OpenAI for the TTFT improvements, esp. on queries with 100k+ tokens.

@levie I use this a lot with coding agents, but I haven’t seen any other UX paradigms outside of deep research-like results, which doesn’t feel like a great comp. Any other software that’s starting to do this?

Almost everything about how we’re building enterprise software products is changing right now. For years when you built SaaS products, your entire design focus was how a user would interact with the system to accomplish their task, by themselves or collaboratively. Now many of the core design challenges are more about how the user will work with AI Agents to do this task. This turns the questions into how the user will setup, deploy, orchestrate, or provide context to AI Agents to execute work, and then review and incorporate their work after. Does this happen in an existing UI? Do you do this through chat? Is it through task list or queue? Is it through a workflow builder? Do you describe it as an Agent or just productive the specific outcome the customer wants? You can see the early differences playing out just in the AI coding space right now, where there’s a debate between the IDE, terminal, web UI, or just using slack. In all cases, one thing that seems to be clear is in many ways, to the user, software directionally gets simpler. The nobs, toggles, switches, and components needed for people to execute tasks are less necessary in a world of AI Agents. The APIs to these capabilities still matter for the AI Agents to use (so they don’t go away), but they’re primarily leveraged in the background. And when they eventually show up for the user, it’s more for advanced use cases, exception handling, or the review process, as opposed to the common activity. We’re in easily the most interesting period of software design that we’ve ever been in. We have to design for users as well as autonomous agents at the same time. And we’re only in the beginning phases of what that looks like.

@sqs I find it's ~fine at low info density prose, where I'm communicating 1-2 simple concepts in a facade of professionalism. But when it's nuanced, the robot can't know the subtleties, and I find that the act of writing helps me understand what I want to say.

LLMs are still terrible at writing prose. There's so much information implicitly encoded in /how/ something is written, and that information is lost when LLMs write for you.