Frosty

[Frosty @frosty_buk]

@MatthewGunnin @supermemory @memdotai do you find any issues when it comes to compaction with a small model that you're using? I tried using deepseek v4 for compaction with hindsight but it started costing a pretty penny.

[Frosty @frosty_buk]

@VictorTaelin This is genius

[ShibaDad8 @shibadad8]

If the Democratic Party was anything close to Zohran, we would be in for generational change in this country.

[Frosty @frosty_buk]

@Saboo_Shubham_ @unwind_ai_ this is pretty sweet. Is it launching codex app server in terminal for each goal session? does it track on kanban throughout the goal or after the goal is finished?

Would love to try the skill if you’re planning to open source!

[Frosty @frosty_buk]

@bcherny @EricBuess @kaiofreitas Is this the same version of loop that was released to get people to stop using Ralph-Wiggum loop

[Frosty @frosty_buk]

@Teknium you cooked with the /goal command on Hermes. Thank you!

[Frosty @frosty_buk]

@RnaudBertrand When you actually use deepseek instead of using benchmarks as your truth, you start to realize these models aren’t nearly as capable. Great for subagents with small niche use cases, but the moment you throw complexity, they start to break down

[Frosty @frosty_buk]

@raju_datla1 @quxiaoyin Qwen 3.6 has been pretty good as a coding agent for me. Chinese models aren’t great orchestrator agents though

[Frosty @frosty_buk]

@blueemi99 You’re a legend

[The Ten Thousand Man @slime_machine]

lot of insane shit has happened in the past year but something i still regularly think about is when charlie kirk was shot in the neck animal style and died instantly benjamin netanyahu went on tv and said he didn’t do it

[Frosty @frosty_buk]

@Teknium would be helpful to have the agent give a short message on what it’s going to do before it starts all of its tool calls or skill searches.

[Frosty @frosty_buk]

@nummanali how is this different from wterm?

[Frosty @frosty_buk]

@iamtrask any kind of orchestration sucks without a large model. qwen3.6 is decent at coding but once you’re Claude pilled it’s hard to see local as valuable outside of cost.

[Frosty @frosty_buk]

@ryanprasad_ai @Teknium @marcosnils @NousResearch Curious how that plan is working for you? How are the usage limits so far?

[Frosty @frosty_buk]

@anamhira Game changer, thanks for making it open source

[Frosty @frosty_buk]

@ASvanevik GLM 5.1 is great as an orchestrator for Hermes for me.

I’m using it to write code now too, with a review from Claude code CLI after it’s done but the code quality is kinda trash, never works on first try.

[Frosty @frosty_buk]

@elkelk Such a retarded take. Changes can actually happen BECAUSE he has influence.

[Frosty @frosty_buk]

@ChrisZazula @vectro @Teknium You can adjust context compaction. 2.7 has a million token context window. Usually hallucinates after ~250k though but not sure if Hermes agent harness would help with that

[Frosty @frosty_buk]

@shaunralston Considering dropping my Claude max plan and downgrading to pro just to have conversational Claude and use Codex pro plan for all dev work.

[Frosty @frosty_buk]

@mckaywrigley Interested!