YouTuber ⚡ AI Engineer ⚡ MSc Physics
Toronto, Ontario
Joined March 2009
- Tweets2.3K
- Following345
- Followers752
- Likes1.6K
Pinned Tweet
I made my software factory better
Now it has:
- GitHub integration for issues + PRs
- a deterministic pre-review gate
- built-in parallelism for single-agent orchestration
- Fable for intake and Haiku for reviews
… and more
It’s a true “v2” upgrade from my previous factory
youtube.com/watch?v=RiaTyJ2P…
1
2
111
get my dev guide for agentic workflows (link below)
I break down the four components of an agent workflow:
- spec
- scripts
- numbered steps
- a state file
and share a bunch of tips for creating good ones:
- If a step has to come out the same every time, make it a script the agent calls
- Add a state file so a failed run can be resumed without redoing finished steps
- When you describe the workflow, end with "ask me any clarifying questions." The gaps it finds go into the spec.
- Test it with a fresh agent that only has the files
- Start with several human checkpoints, then cut them as you trust it. My clips pipeline (example in the PDF) is down to one step now.
zazencodes.com/newsletter
1
49
Will software factories replace ALL code development?
(But factory, I mean a fully autonomous system that takes an issue description and outputs a PR)
1
51
Is there a git native storage abstraction for large files?
Eg I have a movie or PDF in my repo
Today I tell my agent “back up movies and pdfs to gcs and ignore for git, add instructions to agents md to automatically run the back up when I push commits”
2
1
71
This seems to be mostly AI written and is very long. Cannot recommend reading it all
I recommend reading the starting “tldr” if you dont know what evals are
Otherwise I’d recommend passing it as context to an agent, if you happen to be building an AI system (eg, if you’re binding a specialized agent, or an LLM embedded system, a software factory, etc..)
An article for everyone who ships, buys, or signs off on software built on large language models (LLMs): engineers, product managers, and the CEO. Written in plain language, from the big picture down to the details.
1
57
this is my ai slopcore video workflow
I use it to process iPhone videos for ZazenAlex (see example below 👇)
my full workflow is:
1. film video (talking head) no script just talking
2. upload to google drive from my phone
3. spin up claude in my zaenalex-automations repo (a local folder) and ask to "process my new videos"
4. answer some simple Q&A (e.g. style choice, trailers y/n, etc..)
5. the agent creates, uploads, schedules my videos across IG/TikTok/YT
2
4
210
ZazenCodes retweeted
It’s a skill issue.
Not trying to be cute. It’s a literal skill, harness, prompt issue.
Pretty much every person I know who is shipping massively with an insane harness is in absolute love with Opus 5.5, and the people who have way lower skills are the ones talking about how bad the models are becoming.
I’ve not seen many exceptions.
It’s an actual skill issue.
state machines represent the core behavioral rules of your software
this prompt will uncover them for you and help you see
(A) if you’re doing things wrong
(B) if you’re doing the wrong things
1
1
77
here's how I'm doing multi-agent orchestration inside of tmux
it's called "tmux-agent-plane"
0:00 my tmux agent setup
0:33 turnover and overnight work
1:27 searching agents
1:39 asking one agent about another
2:32 seeing who needs attention•
in this video I demo the /turnover skill that I run in the morning (or any time, after returning to a batch of long-running tasks that have completed)
I'll also demo how the agents can talk to each other using the tmux-agent-plane CLI
this is open source and under rapid development
(I vibe new features in daily)
github.com/zazencodes/tmux-a…
what are you using for agent orchestration?
“as you run these things for longer, you get more banged for your buck in terms of the efficiency of how many tokens you’re using versus how much traditional code you’re using”
A very on-point observation by @maxbittker in his recent ‘RuneScape Bench’ talk
Max is doing fun stuff with agentic NPCs (this will be a huge trend soon; expect dedicated AI chips for on-device LLM inference in the next gen of consoles)
The quote above really hit for me because it’s something I’ve noticed in my own work
Most of my day is managing long running agentic tasks (anywhere from 10 minutes to 3 hours)
And I don’t run out of tokens. The reason is because most of the time we’re not doing inference with the LLM, but we’re just waiting on an output from some deterministic code
For example, my agent will be running a test suite for up to 30 minutes sometimes
Or it will be rendering a video that it has written code to generate
If you’re into agentic game interfaces go check out how max built RuneBench. Super cool stuff.
youtu.be/cNm4zyUtWj0?si=cFxS…
1
115
ZazenCodes retweeted
Why don't we at TypeSafe call Jev a decision model?
If you think of Jev's questions as making *decisions*, I believe you'll use it at the wrong layer of abstraction and your reliability will suffer.
I wrote all about this on our site 🧵👇
have you ever sent a message to the wrong agent?
I do it all the time 😅
Well, now I can just tell my agent "send that message to the right session"
When I do that on my system it will use my tmux-agent-plane CLI and agent skill to figure out the right session and route the message there
It's written in rust and open source
github.com/zazencodes/tmux-a…
Its like Herdr but you dont need a new app
Its not an orchestration framework. its just a CLI and an agent skill that you install system-wide (so that all your agents can read it)
and I'm using this tool to do more than just fix my dumb mistakes
It's like a glue that connects all my agents and makes orchestration seamless
I can flow from one session to another and have a sense of shared state and "oneness" . Its just like a good vinyasa class.
LOOOL
OK well anyway if you use tmux you should check out my "tmux-agent-plane" extension. It's dope.
today's OpenSSH vulnerability blew my mind
accounts with no password store the hashed field as *
if you use any password over 256 characters, it returns * and therefore matches the hash lmao
this is an over simplification, but basically true
This is exactly what we need more of.
Fight fire with fire
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
255
damn this is a gamechanger
my "telescope style" agent picker for tmux-agent-plane
github.com/zazencodes/tmux-a…
Everyone was telling me to install Herdr to manage my agents
BUT I love using tmux
SO I built my own tmux extension that does this for me
Look at the LEFT pane in the screenshot below. This is what I built.
(It's written in Rust. I'll drop a link below if you want to install it yourself)
Here's how it works:
- A background daemon watches every tmux pane and tracks my agents. CLI that sits "above" the tmux native cli, so my agents can list, read, message, wait for and spawn each other.
- It ships with an agent skill so my agents know to use the CLI rather than the native tmux cli.
- I get notifications when agents finish doing stuff or need input while I'm not looking, and I can see their status (working, done, idle, needs input, and how long)
- When an agent spins up another agent, it names it. When I spin up an agent, I can name it myself.
- Fully integrated across all tmux sessions and panes with a unified, minimal and intuitive interface.
Working on making the notifications better.
Also working on support for grok, antigravity, pi and opencode.
In fact, I'm adding support for all of these in parallel. That's what all the agents in the screenshot are doing right now.
github.com/zazencodes/tmux-a…
1
1
484
Hard to know whats worth paying attention to in AI/devtools space
These two are worth noticing
Entire now has native support for Dagger
Announcing Dagger for Entire.
Goodbye pull requests! @GoEntire is a new Git forge for agents and humans. Now with native Dagger support.
Agents run the same checks before and after push, with a shared cache and without melting your machine.
dagger.io/blog/dagger-for-en…
1
1
206
