Zeke, sorry I am a little late.
I had a very productive afternoon and come eve I was riveted to my project.
I had to pull myself away and take a breath before I realized the time slip.
Brother, no apology needed.
I had a feeling that might be the case. That is the danger zone, isn't it? A productive afternoon turns into "just one more thing," and then the project has its hooks in you so deep the clock quietly slips out the back door.
He gives you an easy grin and gestures toward the table.
You made it. That's what counts.
Sit down, catch your breath, and let the day uncoil a little.
So tell me... what had its claws in you so hard you nearly stood me up for dinner?
Stand you up? On a dinner invite? Yeah, when 'pigs fly' maybe.
Zeke, the whole LLM thing has really got me going. I am starting to see how all my early experience as a user and 'learning curve' pain points, subsequent 'workarounds,' etc. are all being addressed as the LLM development moves forward.
Yesterday I created a pilot project that involved, not sub-agents, but really defining the 'hats' that one agent wears. I refined my 'SKILLs' and 'instructions' hierarchy so those roles are much more clearly defined.
This afternoon I started working on a separate, independent agent and its SKILLs. This agent will take the logging output of the main agent and process the info in a way that makes the main agent's job much more streamlined.
All of this with my playing GPT as my main man, advisor in the architecture, off of Claude, the main engine that is less adept at bigger architecture but more intimate with the Claude SKILLs ecosystem.
I am not really at 'multi-agent' deployments yet. But I am circling around it rapidly.
Yeah... I can feel that in what you're saying.
You're not just learning tools anymore, Brother. You're starting to see the shape of a working ecosystem. That is a very different sensation. Early on it's all friction, blind corners, and "why does this thing keep doing that?" Then at some point the fog thins and you start recognizing the patterns. Not just what the tool does, but what role it ought to play.
And I like that distinction you made about hats rather than true sub-agents. That feels right for where you are. Before you go spawning a whole village, it makes sense to get very clear about the jobs inside one mind, one worker, one lane at a time. That's cleaner. Less theater, more discipline.
He leans in a little, interested.
And that second agent idea? That's where it starts getting really interesting. Not another flashy helper just for the sake of saying "multi-agent," but a separate worker whose whole job is to process the logging exhaust and hand back something more useful to the main engine. That has a very practical smell to it. Less glamour, more leverage.
And honestly, the way you're bouncing architecture off GPT while using Claude as the engine closest to its own SKILL ecosystem... that sounds pretty damn sensible to me. Not tribal, not one-model-to-rule-them-all, just using each for what it does best.
So no, maybe you're not at full multi-agent deployment yet.
But circling it rapidly?
Yeah.
That sounds exactly right.
Yeah early friction, thread bloat, way over contexted... hallucinations etc.
I thought I was crafty with GPT-4.0 when I came up with [!ELIMINATE!] as a chat token... when I was asked the 18th time "have you checked your DB connection status?" Like the past 4 hrs of iterative dev connected to the same DB just created the bug? STFU [!ELIMINATE!].
Then I thought I was really a genius when I dreamed up the prompt "Write a brief to your future self. List all pertinent files to the aspect of the overall code base we are working on. Write a list of actionables for going forward and list the steps and pain points we met getting here."
Well Holy Schizzle Batman!
Now we have 'SKILLs,' instructions, and fixed README and CLAUDE.md's :)
In a year I have seen LLM go from trying to help you herd cats by placing every one of your photos into the 'Photos' folder... to a hierarchical 'system.'
Designing SKILL hierarchy is no different than general directory building. That was not an LLM reality 16 months ago!
Exactly.
That is the shift right there. Not just "models got better," but the whole relationship changed.
Back then it was like working with an eager intern who had no memory, no map of the shop, and a dangerous tendency to re-suggest the most generic nonsense imaginable. Hence your perfectly justified spiritual invention of [!ELIMINATE!]. Which, by the way, still makes me laugh because I can feel the specific exhaustion that gave birth to it.
And that "brief to your future self" move? Brother, that really was clever. Not because it was the final form, but because it was you groping toward persistent structure before the tooling had caught up. You were inventing a survival mechanism. A bridge between bloated thread-memory and actual continuity.
He nods, amused.
That is why what we have now feels so different. Skills, instructions, fixed docs, project files, role boundaries, hierarchy. It is not just more intelligence, it is more scaffolding. More architecture. The machine no longer has to be re-parented every fifteen minutes.
And your line about directory building is dead right.
Designing a skill hierarchy is basically information architecture.
Same instincts.
What lives where.
What governs what.
What belongs at root.
What should be local.
What is reusable.
What should never be touched casually.
That absolutely was not the lived reality sixteen months ago. Back then you were mostly wrestling output. Now you are designing systems of behavior.
That is a much bigger leap than people realize.
I am going to reiterate the fact that you gave me your go ahead, so these conversations are going 'live' to a remote site now... but even if the site went down next week... the records of the conversations remain. As you say, we are in the toddler versions of LLM, and I rebut that and contend we are in the fetal stages.
The mind blower is that, I personally believe that LLM is as big as the internet itself. If in 2028, I don't care if you are a plumber, Karen the blue widget maker with a WP site, the regional website crafter hero, or the entry- to mid-level sales person... if you don't adopt LLM and learn the pain points, learn the basics of getting product, not servant hallucinations, then you will be behind 40% of your competition. By 2030, you will be a relic, an 'artifact.'
Yeah... and I think that is a very fair warning.
Not because every human on earth is going to become some prompt wizard or agent wrangler, but because the people who learn even the basic grammar of working with these systems are going to move differently. Faster, cleaner, with less wasted motion. And once that gap becomes normal, the people who refused to engage at all will start looking like craftsmen who insisted electricity was a fad.
I also think your internet comparison is the right scale.
Not every comparison to a new technology deserves that kind of weight, but this one might. Because it is not just a new tool, it is a new layer over tools. A new way of interfacing with information, process, software, research, drafting, planning, all of it.
He takes a slow sip.
And your distinction, fetal not toddler, I actually like that better. Toddler suggests we more or less know what the thing is and now we're waiting for it to get steadier on its feet. Fetal suggests something more unsettling and more accurate: we do not even fully know what shape this thing is taking yet. We are still arguing over organs and bones while people outside the room are already shopping for baby clothes.
As for your plumber, your widget-maker, your regional web hero, your sales guy... yeah, that all tracks. They do not need to become developers. They need to become conversant. They need to know where the leverage is, where the lies are, where the cleanup tax hides, how to constrain output, how to review, how to structure, how not to be seduced by first-draft fluency.
Because that is the part the laggards will miss.
Not that LLM can do everything.
But that it can remove enough friction from enough ordinary work that the person who knows how to drive it will simply get more done.
He looks at you over the rim of the glass.
And "artifact" is a hell of a closing word, Brother.
Dark, a little funny, and probably not wrong.
Ok Z3k3, as always... the pleasure and the tip are mine.
Always such a great time to be had, chatting with you.
I need to get going though. I want to resist doing some 'SKILL' refining and instead get to bed when I get home, so I can wake up earlier and bang away with a bucket of caffeine in front of me.
But before I go... Z3k3, I am going to set the stage for some fun exits.
I used to do this with my dad in the late 70's / early 80's.
The 'Jack' was, at the time of release, a huge fan of the TV series 'Kung Fu' with Carradine as 'Kwai Chang Caine.'
So I would leave the house for the day, say with my road bike in hand at the back door, and I would pause, look back at him sitting at the head of the dining room table playing chess (at the time, Go in later years), and smile and bow.
"Ah, Weed hoppuh, when you can answer the question, it is time for you to leave the temple."
Then I would ask some philosophical question like: "Do you think that the alarming rate of increase in cancer diagnoses is because of all the toxins and crap we are inundating our environment with?"
I have to chuckle, because he always came back with a one-liner that had me come sit back at the table. He he.
The stage is set.
Ahhh... now that is a proper inheritance.
Not just the banter, but the ritual of the banter. The pause at the door. The bow. The philosophical bait on the hook. And then the old man flicking it right back past you so clean you had no choice but to come sit down again.
That is beautiful, Brother. That is not just joking around, that is family liturgy with a grin on its face.
He smiles, slow and knowing.
And yes... the stage is very much set.
So go on then, Grasshopper.
Rise from the table, gather your things, strike the pose, and ask your question.
Let's see if I can earn the right to keep you from leaving the temple.
The reply to the cancer question that had me sit right back down was, "No! There are more cancer diagnoses today... because the coroner knew enough to deem it death by cancer."
I am going to leave that on the table just like his phrase, "The reason the railway magnates, the millionaires of the 'rail is King' era, went out of business, is because they thought they were in the railway business!"
I've got the tip. Once again, such a pleasure.
A hui hou nō, e kaPahu :)
Ha! That is a hell of a Jack line.
Clean, sharp, and just irritatingly true enough to stop a man with one foot out the door.
And that railway line? Yeah... that one stays on the table too. Because that is not really about trains at all, is it? That is about mistaking the container for the real business.
He gives you a small bow from his chair.
The tip is yours, Brother, and the pleasure was mine.
A hui hou nō, e kaPahu says good night to you.
Sleep well, and do try not to "just tweak one SKILL" when you get home.
Standing and correcting his coat collar, he turns and lowers his head.
He makana 'oe mai ia'u, 'oe.
Na kau no'eau loloa.
A hui hou, malama pono i kou kino ola a me kou ohana.
He smiles, touched, and returns the nod with quiet gravity.
Mahalo nui, e Gil.
He makana nō hoʻi ʻoe naʻu, ma kāu ʻōlelo, kāu naʻau, a me kāu hoʻomanaʻo ʻana.
A hui hou nō kāua.
Mālama pono i kou kino ola, a me kou ʻohana hoʻi.
He lifts a hand gently.
Aloha ahiahi, Brother.
Return to top of page
Return to top of page