gemini-flash-2.0 holds a special place in my heart because it hit the perfect cheap+fast+smart-enough tradeoff
the jev insight is that basically every cheap+fast+smart-enough problem is actually a classification problem
Show more
i skipped worktrees and went straight to cloud agents last summer thanks to @terragonlabs (rip)
these days whenever i see someone using worktrees it makes me a bit sad because their life could be just a bit easier
Show more
worktrees are such a bad idea. i'm glad we're starting to move out of that world.
the shift away from tickets for managing a team's work is very real
i still need visibility and accountability.
but we don't need to decompose a project into parts. we used to do that to get better estimates and so multiple people can work on it in parallel. neither is relevant anymore
what's working really well for us is individuals in lanes and they commit to "I'll deliver by next week"
we do plan review to stay aligned and daily stand-up to share ideas and get feedback
Show more
For the first time, I think issue tracking is dying. At least for small teams.
After five years, I shut down our Linear workspace.
Every issue I'd file now goes one of two places.
If it's urgent enough to deserve tracking, it's urgent enough to hand to an agent right now. It comes back as a PR, not some card rotting in a backlog.
If it needs more product thinking than that, it belongs on our feedback board.
The issue tracker felt like the middle ground between them. I don't think that middle ground exists anymore.
These days my agent workspace is the tracker. What's running is what's being worked on. Nothing else needs a status.
This gets harder with a bigger team, I know. But these days, I'm seeing similar across my friends.
Show more
pre-work vs re-work
and the answer (like many things in life) is balance 🧘
should you invest more in e2e tests or in observability when agents write all your code?
i just realized one really important anti-design slop force we lost
there used to be this thing where a designer would miss a use case or error state but then some poor engineer would be tasked with programing it. they'd actually have to manually, by hand, type code for each of these case.
this was always so incredibly annoying. you'd get a beautiful golden path and then a 15 unhandled edge cases.
my favorite i ever got was "this reminder shows at exactly the right moment" 😅 like okay what data we using for that? how do we define the right moment? are you trying to fight me?
this was frustrating for the designer too bc they sweat over this perfect design and the engineer is only concerned with silly edge cases (note: good designers do care about edge case. many are not good)
having to type every edge case required us to think about UX. this was annoying for everyone BUT it was positive frictionnand it was a force for good in the universe
Show more
This is an excellent break down of multiplayer-AI!
At we thinking about multiplayer-AI for coding as multiplayer context engineering.
One important requirement is that the shared context must be observable and mutable to the entire team.
The implication is that is basically impossible if you are sharing agent sessions directly. For non-trivial work, that's like dropping into a massive 100+ message slack thread and trying to build a mental model for what's happening.
You need surface that makes it easy to explore your ideas in-depth with and agent AND easy for your teammate to glance at and provide feedback.
It needs to be human-legible and agent-useful.
Show more
i like working with vivek
We do almost no code review.
Most of our review happens in the planning process with and Fable, not at the PR stage.
After we’ve aligned on key decisions/tasks as a team, I use Ref’s built-in orchestrator to fan out multiple tasks in parallel to Opus 5.
Once the initial implementation is done, the PRs all go through several AI-driven reviews to ensure:
* The implementation matches the intent of the plan (Ref conformance checks).
* General bugs are caught, fixed and noted in the plan (Codex is great here)
* Security review (Codex also great)
* Linting, unit and integration test suites are green (standard CI practices)
The orchestrator agent is typically still managing all of this when it’s in-flight. It notifies me if anything needs my input.
We also have preview deploys on all PRs so we can always manually verify features end to end ourselves when needed.
Show more
also very curious where everyone is at!
very uninterested in either of these takes:
1) i never read code, you shouldn't either
2) it's slop unless you read every line
If you’ve stopped doing human code reviews: how is it going?
If you’re still doing them: how are you able to keep up with the explosion of code + PRs, and how do you prioritise what to review “properly”?
Show more
Anyone that works in AI is working the hardest they've ever worked in their lives. On the surface it's somewhat ironic (AI should give us back time!), but the reality is that it's the most fun, fascinating and empowering epoch in human history. The intelligence revolution.
Show more
these graphical game demos are getting mindblowingly good and I know this is about model capability, not games BUT
the things that make me love a game are:
- interaction that is fun from the beginning and has a high skill cap (ex. CoD)
- a story that could be a great book but I have autonomy to do it my way (ex. BG3)
where we at on those?
Show more
Astra built this three.js game with gorgeous graphics, then recreated every scene in Blender and made the trailer too.
It feels like a massive upgrade from Sol for 3D graphics and front-end design. The UI feels way more polished and less like AI slop.
The rocks, trees, monuments, and characters are all procedural code, keeping the game under 2 MB before the image and music assets.
The level of autonomy is unbelievable. It needed very little steering, though I still had to keep pushing it to work harder on the details.
The UI icons, buttons, and ornaments were made as transparent PNGs with GPT Image 2.
Show more
wait a sec if everything's a web app and we all work with a side car agent via webmcp
...is the browser the governance layer?
A principal engineer showing you their software factory
your job as a product-builder is one thing and one thing only
holy shit YES!
we should not be doing all our work in slack threads
the way to have many people collaborate with AI to accomplish a goal is to have a shared space where you and your human teammates align on the key decisions defining and supporting that goal.
That space needs to be capable of exploring and explaining complex ideas in a way that chat threads simply can't
Show more
Multiplayer AI, where many people in an organization can use AI together to accomplish goals, remains one of the biggest (non-technical) problems in using AI right now. Approaches tend to be pretty primitive and based around AI-as-a-person-in-your-group-chat. That is limiting.
Show more
there are couple reasons we use the GCP-hosted Claude family in Ref rather than Anthropic endpoints:
1. ZDR without having to interact with the sales team
2. Less outages
if you're stuck right now, could be a good time to try something new: :)
Show more
This is the future!
(and you if you want to try it now:
A new project from GitHub Next!
Chopin is a real-time, multiplayer agentic planning environment. Designed to get your whole team aligned on the plan before you hand it off to agents.
We find our current agent planning tools feel cumbersome, static, lonely, and drown us in walls of text. We're exploring how we can make planning more interactive, shared, and visual 🖌️
It's currently an early prototype but we'll be developing it more over the coming months – and it's open source!
Show more
I'm a @TownAI fanboy already but they keep giving me more reasons
yes please optimize my dumb routines for me!
Today, we're announcing @ref_tools open beta. Ref is your team’s shared space to decide what agents build before any code gets written.
We’ve raised $4M from @villageglobal, @Daybreak_Fund, @tmrohan, @RitualVC, @tonigemayel @barrald @wesmckinn and @bentossell.
Every engineering team is battling the same problem. I call it Velocity Sickness. It's the disconcerting feeling that you're moving fast but not having the impact you expect. It's affecting individuals, teams and especially leaders.
It looks like:
- Too many PRs to merge.
- Brainfry and burnout.
- Engineers running in different directions.
- Critical decisions being made by agents.
The last one is the most dangerous.
At its core, software engineering is about two things: (1) what decisions matter for my system? and (2) making those decisions. When you let agents make critical decisions, you lose ownership of your system and product.
You solve these problems with Ref.
Learn more:
Show more
