akira

agent harnesses @wearerandomlabs

[akira @iso_morph]

https://t.co/TrLJZg9cck

[akira @iso_morph]

@raealisa This has to be bait

[akira @iso_morph]

@dexhorthy @trq212 My goat you got this

[akira @iso_morph]

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

[Adi @adi_swaminathan]

It's almost like @iso_morph was ahead of the curve. again...

[akira @iso_morph]

@chai_zhenhua Yes

[akira @iso_morph]

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.

[akira @iso_morph]

@cleinliu Astra on high reasoning

[akira @iso_morph]

@Merlin412188240 Really?? Haven’t used dots, and if I was to use a personal agent I’d probably start w grok bot how do you think they compare

[akira @iso_morph]

@andrew7hess Yes actually

[akira @iso_morph]

@4SimonSays Real

[akira @iso_morph]

@SenseNopedOut Yes it verifies and backs up everything it’s changing lmao

[akira @iso_morph]

@bolayer Speed and UX mostly

Fable has historically been super slow in my experience and also previously had a hard time on this task (have not tried 5.2 though!)

So the models I use are Astra and Opus

Astra is fast enough for me

[akira @iso_morph]

@luisguimontmota Not for me I guess. I still have to think through solving it, but astra very very quickly takes what I say and implements it in the quint format

[akira @iso_morph]

@lisi288917 Yes. This is something like a mathematical proof. In fact I can run it through a proof engine I'm just not doing it.

The TLA model checker does brute force search to find traces where rules for the system fail

[akira @iso_morph]

@blurskye_ Need to yes

[akira @iso_morph]

@adityaa_behera ?

[akira @iso_morph]

@jkgorka 6 Astra High

[akira @iso_morph]

@skopinax0 Likely overkill for most work anyone does yeah

[akira @iso_morph]

@peteystruggles Real

[akira @iso_morph]

I think for this task it might also be in token consumption

It took opus much longer time wise to give me a correct result and if we assume a relatively equal rate of burn then for it to be equal it would simply need to take me 2.5 x as long with opus

Which feels approximately true but again I’m going off of vibes since I didn’t time wall clock time or check my initial usage for either

Certainly opus spent more tokens but the cost difference is massive so it’s unclear who comes out on top

[akira @iso_morph]

@HoboJerk Oh yeah cool

I think state machine verification is pretty good

I think what gets weird is concurrency

[akira @iso_morph]

@tv0ff This is hyperbole but it’s not far off from the behavior I see

[akira @iso_morph]

@tv0ff Sure but the difference is being able to do the work vs not