Ossie Irondi

I build and operate workflow software and services for small businesses, with a current focus on dental operations, market research, and product diligence.

[Hugging Models @HuggingModels]

Behind the scenes at Hugging Face headquarter.

*literally*

[Ossie Irondi @AojDevStudi0]

@mattpocockuk The next evolution of grill me for those with short attention spans could be: AI lists all the assumptions upfront then allows you to only drill down to the ones you don’t agree with. These days I’m agreeing with the LLM 83% of the time.

[Ossie Irondi @AojDevStudi0]

@ChrisSaysSo @petergostev The strawberry benchmark is undefeated sir

[Ossie Irondi @AojDevStudi0]

@benthompson Ponytail + caveman + I have ADHD plugins

[Ossie Irondi @AojDevStudi0]

@lydiahallie Force people to read the docs!

[Ossie Irondi @AojDevStudi0]

I had an epic battle with @claudeai yesterday

[Ossie Irondi @AojDevStudi0]

@danielkleach https://t.co/A8cK3HNAbL

[Ossie Irondi @AojDevStudi0]

https://t.co/Skn5Um04aB

[Ossie Irondi @AojDevStudi0]

Opening a prediction market for @IndyDevDan’s Monday video:

40% — Herdr and the rise of agent multiplexers
35% — Moonshot AI’s new Kimi K3
20% — Running GPT-5.6 Sol inside Claude Code
5% — A fourth thing he discovered before the rest of us

Place your bets.

I think I may have finally front-run Dan’s content calendar.

[Ossie Irondi @AojDevStudi0]

The Herdr Fleet skill in action with Kimi K3 being the orchestrator. This should be loads of fun.

[Ossie Irondi @AojDevStudi0]

So I think the first thing I'm going to do Kimi K3 in Claude Code is I am going to audit from top to bottom the LifeOS system and chop down all the bloat I can find.

[Ossie Irondi @AojDevStudi0]

So I built a PR review bot that does PR reviews for me and scores it, kind of like the way Greptile works. And so it's a really overly engineered system that runs at my home lab. So I don't actually look at diffs on the ClaudeCode app or anything like that. I go to my PR review app and from there I can review diffs. And I also made it look like the way the new CodeRabbit feature allows you to review diffs. And also I keep my PRs super super small, super tiny because I use @mattpocockuk 's skills. So all of my slices are super super small, super tiny. Not like I'm having huge PRs I have to review all the time. And most of the time I don't look at a PR because it gets the score that it needs, it gets a 5 out of 5. So if it's a 5 out of 5 I'm not looking at that thing. But if it's anything less than that, then it will require me to kind of step in and take a look at it. So yeah, it takes a while to set up your whole entire system. But if you take the time out to really build it out how you want it, then a lot of the heavy load and the heavy lifting will be handled by your agent systems. But don't be like me though, I an agentic engineering the fuck out of life.

[Ossie Irondi @AojDevStudi0]

1/ I think I’ve found a genuinely useful multi-agent engineering workflow.

Today, I’m away from my desk running errands and attending a wedding.

Meanwhile, five coding agents are implementing issues and reviewing PRs inside one repo.

I’m supervising the operation from my phone.

2/ The foundation is @herdrdev.

Herdr is an agent-aware terminal multiplexer: real terminal panes, persistent sessions, agent status, remote access, and an API the agents can use themselves.

Think tmux, but designed specifically for coding agents.

https://t.co/rxR7i15cEb

3/ I enter the project directory, launch a Herdr workspace for that repo, and start Fable 5 as the lead agent.

Fable then creates and manages five worker panes:

• Claude Code + Opus
• Claude Code + Sonnet
• Pi
• Codex
• Claude X

4/ Claude X is my setup for running GPT-5.6 inside Claude Code.

So this is not five copies of the same agent.

It is a heterogeneous team of models and coding harnesses working inside the same persistent Herdr workspace.

5/ I divided the five workers into two functional groups.

Two agents are referees.

Their only job is to continuously work through the PR review queue: claim a PR, inspect it, validate it against the issue and engineering standards, and determine whether it is ready.

6/ The other three agents are builders.

They select open issues, implement the work, test their changes, and open PRs.

That creates a continuous internal loop:

Builders implement issues → referees review PRs → builders address findings → referees validate again.

7/ Fable 5 sits above both groups as the orchestrator.

It monitors the workers, delegates tasks, keeps the queues moving, and handles coordination across the panes.

I am not manually operating five terminals.

I am managing one lead agent that operates the five terminals.

8/ Fable is also the only agent with a remote-control connection to me.

When the system reaches a decision it cannot safely make alone, Fable pings me through the Claude mobile app.

I answer from my phone, it routes the decision back into the workspace, and the agents continue.

9/ The operating model is:

• 3 builders implementing issues
• 2 referees reviewing PRs
• 1 orchestrator managing the system
• 1 human handling escalations from a phone

All operating inside a persistent @herdrdev workspace tied to the repo.

10/ This is the part of agentic engineering that interests me most.

Not simply launching more agents.

Designing roles, queues, review gates, escalation paths, and control boundaries so those agents can function as an actual engineering system.

And honestly?

Not a bad workflow whatsoever.

[Ossie Irondi @AojDevStudi0]

I just got Kimi K3 running inside Claude Code 🚀
Claude Code → patched CLIProxyAPI → Moonshot API directly. No OpenRouter.
Kimi is my default model, while Opus, Sonnet, and Haiku route to GPT-5.6 Sol, Grok 4.5, and GPT-5.6 Terra.
One interface. Four model backends. WHAT A TIME TO BE ALIVE.

[Ossie Irondi @AojDevStudi0]

@DanielMiessler I know you're working really, really hard. Just want to encourage you for the next release that you're going to have for LifeOS system. Can't wait for it because I do not have the brain capacity to gut my system.

[Peter Yang @petergyang]

Anthropic recently cut Claude Code’s system prompt by 80%.

@trq212 explains why:

“As the models have gotten smarter, they need less direction, fewer constraints, and fewer examples.

The examples are constraining it because now it’s like, ‘Oh, you want things like this example.’ If you remove the examples, it can actually be more free-form.

All this is to say that you want to trim your context [when a new model is released]. The latest models often need more room to run.”

📌 Watch the full episode here: https://t.co/mMrt4D5qU7

[Ossie Irondi @AojDevStudi0]

I hit 163 GitHub contributions in one day after treating every Fable 5 session like it might be the last.

The useful result was not more agents. It was a system: 3 builders, 2 referees, 1 orchestrator, and 1 human on-call.

Plus the full Claudex recipe.
https://t.co/qXIiVIOiwt

[Ossie Irondi @AojDevStudi0]

I’ve been sleeping on @cursor_ai cloud agents @Rasmic you gotta redo that video you didn’t fully explain the sauce! Somebody help me pick up my jaw!

[Ossie Irondi @AojDevStudi0]

@steipete Peter! Stop it! Give us 2 more weeks please!

[Ossie Irondi @AojDevStudi0]

@MandleHu @claudeai Lack of maturity within the typical American customer never ceases to surprise me! They don’t have the same capacity as openAI it’s that simple

[Ossie Irondi @AojDevStudi0]

@rayimbuilds Yes absolutely, turn on remote control with Claude code or if you’re using GPT-5.6-sol as your orchestrator then setup codex mobile and you’re good to go.

[Ossie Irondi @AojDevStudi0]

https://t.co/e1k5nlCXCX