Building agents that automate everything in your company: https://teapawt.net/t.co/hhaxPjaLDL / ycombinator fellowship / top 80 NA FIFA player
Berkeley, CA
Joined October 2012
- Tweets2.6K
- Following223
- Followers403
- Likes1.9K
Pinned Tweet
#throwbackthursday
Back before I built software and my dream was to make it as a pro soccer player
AI this, AI that, whenever I’m back in the south all I want is a good ol’ steak and eggs from @WaffleHouse
You’re getting lapped 15x over and you don’t even know it
Everyone knows speed is the #1 pillar of startup / business success. Speed of implementation = speed of experimentation = finding pmf faster
This is a PSA that you could be shipping much much faster
Man, if you don’t have agents teams working over days, you’re really missing out on the productivity / alpha you can have.
One project’s been working for me for 3 days, wouldn’t have been able to ship this previously unless I put 100% focus on it
27 PR’s shipped, 2 to go 😙
I had to buy my own server to support the increased shipping speed we had.
We’re shipping 13x faster than in May. 5 months, 13x speedup.
The world isn’t ready for all the compute we need for agents
This week we didn’t see an uptick in shipping velocity. We stalled due to CI capacity issues with our provider, and to make things worse, even our cloud providers have run out of capacity to host our runners. VCs are getting excited about funding faster/cheaper runners while the reality is that constantly optimizing and rewiring CI/CD has become the name of the game. Once you solve the agentic validation and code-review bottlenecks, you still have plenty of engineering and janitorial work to keep the factory humming.
All this talk about PR’s when your amp orb is just a PR with a different skin.
We’re not using PR’s cuz we have other devs review them. It’s because we don’t need to build a new UI layer for the agents to talk to each other.
They post on the PR to record their decisions. They use “ready to review” as a way to know when to do adversarial review.
Saying “you don’t need a PR” is stupid because yes you don’t need a human to review it, but you still need all the UX layers around it, even if you mask it as something else.
Did this with our own PR review.
3 rounds, but instead of asking the human to change the approach, we ask the agent writing the code to rearchitect.
results:
Cap your agent review loops at 3 rounds
Per their writeup, one Claude Code PR ran 50+ rounds of review and fixes overnight, mostly re-flagging old code. After the cap, PRs needing 5+ rounds went from 8.9% to 1.5%
for your REVIEW.md
After 3 rounds, stop and ask the author to rethink the approach
teapawt.net/ClaudeDevs/status/2111…
10000% cope.
Humans are no longer the maintainers of code, agents are. If you’re not treating agents like a senior dev and you’re the PM / team lead, yngmi
Persistent VM’s are an anti pattern and poor / lazy engineering.
The real sandboxes are ephemeral. The workspace lives while the sandbox can be swapped out for any live box. Your context, work, all stays, but the machine running it can be any machine in the world.
Cloud / distributed systems already figured this out a decade ago.
Curious that many agent providers still haven’t.
seems obvious to everyone around me that sandboxes are on their way out. we hear too many stories about reliability, capacity, broken sandboxes that need manual intervention -- plus they are way too expensive for what they provide.
the future for agents? workspaces.
your agent doesn't need a computer. it needs a place to store it's context, and the *tools* to manipulate that context each turn.
we're in the very early stages of the AI era. the simplest version of this pattern is just handing each agent a time-multiplexed VM, since that's the only thing that we have at hand right now.
over the next 12 months, you're going to see a rapid shift away from this model and towards what i call the "magic compute box".
rather than giving an agent a VM, you give it a set of tools (could be "bash", "python", "js", "read_file", "write_file") -- and you send those commands to a magic service in the cloud that manages to run them for you -- using the cheapest technology available (just-bash, wasm, or even Linux if really required) for each command invcation.
your agent doesn't want a computer that it can set a desktop background for, it wants a place where it can fire off an unlimited number of tool calls and have it update the context correctly.
Shipping ~8x faster than in May, ~4x faster than in August, ~2x faster than in September
Don't tell me multi agent / orchestration don't work. You're not doing it right.
If you're not using at least 2 computers, and running 100 agents across your stack, then what are you doing?
This is the way
.@sama says the advice to pick one thing to build and pour all your focus into it no longer applies.
"The canonical advice that was correct two years ago was, 'Pick one, and put all your effort into it, because you can't build all 30 things.'"
"You can now build all 30 things. And you can get people to use them all. And you can work more on the ones that work."
"So I think this mindset of lots of ideas, quick feedback, and being open to a lot of things not working is great for this new era for how people build."
People will say software factories don't exist, and you're wasting tokens by having multi agent setups
But then how do ppl explain this?
Definitely not. That's why you should make your own company workspace with codepress.dev
You own everything. All the code, the workspace. We just host it for you and give you easy defaults.
If you're using the frontier lab workspaces, what are you doing?
