Lmao, we built the first swarm based agent... 8 months ago at this point!
New version of slate is going to lean into this even more
5
673
Definitely on RL hotpath. Astra writes comments directly in the code in an almost note taking way that I've only seen in TLA. In Typescript Astra's code is just minified js basically.
Astra 6 is so completely brutally better than Opus 5.5 at formal verification it's not even funny
I'm using TLA to verify something we will be launching soon, and... man Opus is getting completely fried and Astra is just cruising.
8
1,157
Vibe formalization works, as long as it is entirely agent in the loop.
The main issue is that if the agent is not the one making decisions, then the context is not present for the vibe formalization to be coherent in the same way that code is less likely to be coherent if the agent isnt fully writing it.
This also implies that more autonomy, not less, is better for coherent code. The issue is that it might not be the code you want!
What the
Some news: I’m joining OpenAI to lead product management for Core Products!
I have a lot of respect for OpenAI for showing the world that scaling works, for bringing AI to over a billion people, and for believing great products are core to the mission.
I’m excited to work with @thsottiaux, an incredible PM team, and everyone across OpenAI. Together, we’ll build agents that have your back and your team’s, and keep you in control, so you get more time for the people and work you love.
Lots to learn, and lots to build. Please keep your feedback coming!
1
7
1,110
Not only am I back to agents, slate is getting a massive overhaul
What do you wish your agents could do on their own? And what have you hacked together for them?
Lots of people talking about their software factories
And how much of that swarm scaling is coordination overhead?
I'm finally working on actual agent engineering again.
A few observations:
Seems like everyone has finally caught onto the idea of sliding window compaction.
...
5
18
1,225
Been messing with Opus 5.5 and boy does this model love to overcomplicate the abstractions it is implementing.
So many things can just be written inline very simply
What's interesting is it is reasoning more clearly and directly about patterns it uses
Pretty good model
3
13
1,477
An agent can do all the non-structural parts of systems eng as far as I can tell
This is actually good because you can design systems and think through them and then magically they exist
3
13
1,110
Opus IS significantly faster than Astra xhigh for what its worth
back in the distributed systems bunker for a bit today
Testing opus 5.5 vs Astra
Its interesting. Opus 5.5 feels like an Opus model
Astra is somewhat different in how it approaches its correctness. Opus seems to want to shoot for a full solution that I would like to read, whereas astra is mentioning invariants more. I need to get opus to use its math skills more
1
20
1,906
back in the distributed systems bunker for a bit today
Testing opus 5.5 vs Astra
Its interesting. Opus 5.5 feels like an Opus model
Astra is somewhat different in how it approaches its correctness. Opus seems to want to shoot for a full solution that I would like to read, whereas astra is mentioning invariants more. I need to get opus to use its math skills more
The cluely architecture diagram got one thing right
@Cloudflare workers just scale
Infra problems go down number go up, still not measuring the underlying perf though
Just started running our baseline evals for new slate
I ran a really untuned harness, astra + bash, on TBench 3 and out of 53 runnable tasks, there were 19 failures due to daytona rate limits lmao
Need higher rate limits I guess
1
1
1,330