brumbus

3D Pinball for Windows has obviously already been reverse engineered, and the agent itself speculated that it might not be able to credit all of its insight to active reasoning. but based on what other folks are saying, it works just as well on more difficult stuff.

this is obviously a huge deal. agents are very good at this and REA is an empowering interface. everything is open source now, at least kind of, for a one-time cost at pretty reasonable rates. my pinball experiment cost $3.29.

so ghidra worked better but I gotta say, hopper's UI opening, flashing a bunch of hex and assembly, and then closing quickly while the agent worked was extremely cyberpunk

agent commented that using the REA MCP server was the trick to going fast (otherwise with the CLI you keep expensively restarting ghidra over and over)

that CLM thing just sounds like a good way to invalidate most of your cache every turn imo

lmao codex bedrock hard-codes the auto-permissions model to global.openai.gpt-5.6-luna with no way to change it. at work we can only use us.* endpoints!

opus 5.5 on bedrock is shockingly unreliable. long outages the past 3 days. us.anthropic.claude-opus-5-5 specifically

“they stole” ok then why do you still have it

I use AI an immense amount, every day, and I still do not understand at all how or why people have dozens of persistent agents constantly working. How many emails do you receive a day? How many projects have you amassed? Don’t you finish tasks? idgi. Maybe this is related to my not keeping tabs open

update: this is going very well

brumbus@brumbus.bsky.social

joining the "well I guess I will just maintain a prime-agent fork" club

absolutely crushed our AI cost by pushing folks toward openai models for subagents. we let folks use opus or astra as they prefer, but insist on sol and luna subagents. we’ve gone from a nearly exponential trend to nearly flat, despite headcount growth

Doomsday, or the 120 days of Marvel

bedrock's reliability issues are killing me

i'm gonna cut your hair, and put it in a bowl

a mood house?

joining the "well I guess I will just maintain a prime-agent fork" club

Sol 6 actually beats Astra on our internal company benchmark

barely on raw score, massively on cost

the economics of testing code is changing, imo. 2 minutes of testing to 60 minutes of coding makes total sense, but 2 minutes of testing to 30 seconds of coding is different. testing is more important than ever as we take our hands off the wheel, but which tests and when is important.

(it was already important when we were still doing it manually, but it's moreso now)

it's super neat that sibling agents can communicate easily in prime-agent but I don't feel like it actually helps. heavy coordination slows things down a lot, plus it's expensive. each agent message is two calls.

I feel like I'm hacking the gibson when it launches 10 subagents and half of those launch their own, but is it efficient? I don't think it would have taken 1 astra 90 minutes to do what I asked of it. time isn't the only axis to care about, but it's one of them

and this is with the "all queued messages arrive at once" mode. don't get me started on the default one-at-a-time

Woo-hoo! Look at those tokens fly!

how can Jev play Doom when I can't even get it to play tic tac toe very well?

I think it's that Doom play can look proficient even when mildly sloppy, whereas rigid discrete games like tic-tac-toe and chess aggressively punish even a single bad move.