Building AI harnesses
Joined November 2022
- Tweets 385
- Following 295
- Followers 1,215
- Likes 324
97 Photos and videos
I eventually settled on the following approach:
β’ Install and configure Herdr (herdr.dev/) on the host machine.
Think of Herdr as a tmux-like session manager for AI agents, allowing agent sessions to persist independently of your client connection.
π
1
123
This keeps agent work isolated and makes it easy to reconnect to the exact session you were using, regardless of which device you're connecting from.
When I want to inspect or edit code, I connect to the same host using VS Code's Remote SSH extension.
π
1
41
This gives me direct access to the project files while Claude Code, Codex, or Pi continue running uninterrupted on the host machine.
37
Recently, I've been working on a better setup for using coding agents like Codex, Claude Code, and Pi across multiple devices.
My goal was to run the agents on a dedicated host machine while controlling them from my laptop, so I could work from anywhere in the house. π§΅
3
1
132
Trying to build my own opinionated agent harness orchestration.
->Building upon the pi agent sdk.
-> Has custom tools
-> web search/fetch capabilities
-> Designβs mock
etcβ¦.
Will be writing about it soon with a demo π
1
170
Every AI browser tool I've used is slow, burns through tokens, and spins up its own Chrome, making me log in again every single time.
So I built a browser harness for the pi coding agent package that fixes all of that, and gives the agent way better access to the page. π
2
240
1
79
For those who do not know what pi coding agent is : pi.dev/
92
Finally tested the full video streaming pipeline π
Supports VOD Live.
For non-video folks:
VOD: Upload video β auto-compressed and multiple qualities β direct streaming links.
Live: Send a feed β processed & streamed to viewers in ~ real time.
2
4
211
More features coming soon:
β’ On-demand thumbnail generation
β’ Base file index generation
β’ And moreβ¦
62
Current features
β’ Auto scaling
β’ Smart bitrate ladder generation
β’ GPU auto-detection hardware acceleration
β’ Highly configurable
β’ Live streaming (WebRTC, RTMP, SRT) DVR playback
69
Finally done with the object storage that I was building for learning purpose.
It has:
β’ SHA256-based deduplication
β’ Resumable chunked uploads
β’ Signed, expiring URLs
β’ Password-protected shares
β’ Separate garbage collector
Link : stash.aksdev.me/
Please dont flood;)
3
141
Today's work
Download flow: Opening a shared link now generates a download URL when the user clicks Download.
Password protection: Users can set a password for uploaded files.
Frontend: Vibe-coded it⦠because why not ;)
I am actually very fascinated by object storage services and would be studying more in detail about it. Will also try to make a toy version of it.
144
I wanted to track the chunks being uploaded as part of the multipart upload feature, so hereβs how I implemented it. π
I am actually very fascinated by object storage services and would be studying more in detail about it. Will also try to make a toy version of it.
1
167
Sets does not just come with integer, it comes with hash buckets, pointers, and metadata. Now multiply that by 1 million entries and memory usage balloons fast. So what would be a better approach?
Use a bitmap (bitset). Here's why .
1
126
Instead of storing objects, you store bits. Each chunk index maps to a single bit:
0 -> chunk not uploaded
1 -> chunk uploaded
And for 1000 chunks we would only use 1000/8 or (1000 7)/8 -> 125 bytes
and for 1000 uploads this will be 125 x 1000=125 kb of data. Efficient right?
90