正在加载视频...
视频加载失败
I just recorded a deep dive into Pi's architecture. It's a minimalist AI coding agent with a beautiful design. Here's how it actually works 👇
51 条评论

The video is out on YouTube with chapters and bonus section on skills at the end 🤓

need part 2 🙌

you got it 🫡

Is this video on YouTube too? I hate watching this in x app directly

well it was scheduled for monday, but i might as well publish it today since this post got traction. that'll give you the weekend to explore pi. will post as soon as it's up

This was awesome i agree the parenting thing is genius (Rebuilding this in our corp custom so super useful!)

glad this was useful!!

What's your take on how Pi handles context switching between different coding tasks? Most agents I've tested fall apart when you pivot mid-conversation from debugging to architecture discussions.

i think pi does a great job on this because of its /tree session history. you can finish the planning stage, fork the conversation, have the planning session summarized, and switch to building, debugging, etc. it's a great feature

Thank you.

Good stuff. But, dude, the handwriting... wow.

hah is that good or bad hand writing??

Kept me on my toes, for sure.

you are going to upload this video to youtube too, right? I want to watch this with NotebookLLM.

it's out!

this is so good and might be the thing to get me off of Claude Code. Looking forward to the video being posted on YouTube!

the grass really is greener on this side 🤓 jk, i'm a huge fan of claude code, but recently i feel that the harness is too bloated

it's out on youtube btw!

fantastic - thank you!

Amazing video, this is a Harness Masterclass actually

i appreciate it!

Is there a YouTube link?

it'll be out on Monday!

well, ended up posting it today. here it is:

Thanks!

Thx. Very high educational quality.

i appreciate it 🤓

super helpful, alejandro .. you were about to move to 'skills' where the video ended? .. that's intentional?

heh not intentional. full video out on youtube now though with the skills part:

Silly to ask, but could you please tell what you used to record and edit this video?

the one and only @screenstudio

Do you have this video on Spotify?

not on spotify, but coming out on youtube over the next few days 🤓

it's on youtube now btw

This is going to be so helpful pi so so good and underated agent

agree

Yes

parenting is elegant but does it handle dirty failures? network partitions and orphaned subagents. that's the gap between demo and shipped.

🎮

Excellent. As always enjoyed this video tutorial from you 👍👍

are the tool calls included in the chat history?

yes they are :)

help me better understand the pi concept of trees in its session management. What happens when you fork a new tree branch from an earlier message? Will that basically create a new session, with all the messages so far in its context? Can you then work of both tree branches independently? Or is it more like a rollback thing where it will undo any tool calls that might have happened before? Also, could you explain what the effect on kv cache would be when creating a new tree branch? Would the model need to keep two separate caches or can the first common part of the cache be reused (is that perhaps prefill cache purpose?) note, im an amateur, be gentle with me

pi is the most underrated example of what happens when ai meets software dev simplicity

imo 'minimalist coding agent' just means it hasn't met a legacy codebase yet

Brilliant video - I'm using Hermes and very interested in using pi so that I'm more in control 😎

great breakdown. the architecture decisions around agent memory + tool routing are where the real engineering is now — the model call is almost the easy part. the orchestration layer is the product.

I really want to use Pi, but I cannot afford the tokens at the API pricing.

you can use your codex sub with pi, though! i pair a codex sub with hf inference providers and it's REALLY good. you can use open models that are 20x cheaper than closed models on pay-as-you-go api. and for heavy lifting, you just switch to gpt-5.5. works like a charm

@ydennisy_ Can you use codex, Claude and Gemini CLI through the subscription with PI Alejandro?

@ydennisy_ codex YES! claude and gemini are a bit less friendly 🥲
