finally found time to make a new video! this time i'm sharing a slightly more advanced agentic engineering session, focused on high throughput multi-tasking hope you find helpful!

brb kun dropped

Arrrgh 🏴‍☠️

🍻

After watching this video, started using grok 4.5 as firstmate and i like how its straight and to the point. I was using Fable low until now and even in 20x that runs out pretty fast.

Thanks! will have a look Btw what is your current fav model for first mate? I find the current setup asks a lot from the model, and the steering hooks are pretty heavy, which dilutes the attention of the model, and burns a lot of tokens Had success with Fable 5.1 and Kimi K3 but it's not practical as it can chew through your entire quota fairly quickly. Any plans to make it lighter token-wise? I feel like some processes currently done by the LLM could be done deterministically in code? And thanks for all the work!

my default firstmate model is grok 4.5 right now. depending on quota i also use opus 4.8 sometimes

I greatly appreciate your kind sharing!

Your Firstmate has been instrumental to better work-life balance. Love it!

Love it! Lavish don't get enough credit but I think it's one I can't live without!

Damn, i didn’t know u had a yt channel. Just subscribed!

Kun is the goat.

always useful and inspiring ur content bro! look forward to watch it!

This is gold @kunchenguid ! Thank you! I also love your honesty and transparency about everything!

Exciting! Will watch it. 🙏

Excited for this one

BREAKING NEWS: New episode from Kun !

🤣 it’s been too long

Boom, can‘t wait so see

I really appreciate people like you . Thank you

Firstmate is the best! it change the way I am working. For me it works well only with Fable 5.1

Love your instructional flow, easy to follow along. Thank you!

刚好最近在做智能体多任务调度，正卡瓶颈呢，这个内容来得太及时，马上蹲视频学习。

😁

Thanks.

curious to know how much do you spend on tokens each month

Thanks 🐐

Excited!

Kun Chen!

high throughput agentic setups are great until shared context starts duplicating across parallel tasks and the token bill looks nothing like your single-task runs

Thank you Kun for blessing me with your elite knowledge

Keeping track of several tasks is the part I’d want to learn. Starting more agents seems like the easy bit.

Multi-tasking is where agent UX gets interesting.

Nice deep dive — readable skim here if anyone wants the text version:

Tried high throughput multi-tasking with agents. Agent throughput went way up. The review queue still has one worker, and it's me.

Awesome, do you plan to ‘abstract’ messaging layer? I’m struggling with durability on messages and I’m thinking relying in some sort of queue.