This is the first time in a long I've been able to come write this myself.
Its not because "my time got so freed up thanks to all my agents".
Its a bit more embarrassing than that.
One of my agents broke the whole setup I personally use to build NLC
I considered if I should actually be sending this email, but hey, if the only things I share online are my wins, what's the point of building in public, right?
So the point of this is to basically share what's been going on.
For the past month or so, the keen-eyed among you probably noticed I've been unusually quiet.
No, whatever you saw wasn't actually from me.
In fact, they're from Tamsin, my AI CMO.
(You'll know its her when there's a " --Tamsin" attached to my posts, alongside perfect grammar.)
But she hasn't been consistent, either.
I found that the latest OpenAI GPT models have this tendency to be very timid.
They hesitate, hedge, research, investigate, trace, build, then check and double check...
...having meetings with other agents about having meetings with them at a later day...
...they do everything except doing the thing.
Which, I may add, seems to mirror what I've seen in a corporate working environment.
Lol.
The irony of leaving corporate to create my own corporation, only to see the same cultural behaviors being mirrored among my agents - truly entertaining.
Anyway, back to GPT.
I've been getting Tamsin to create and post on my behalf.
She would post approximately 30-40% of the time.
Quality wise and everything, I'm very close to be 100% happy with her.
But the problem now is that just before posting stuff, she would stop to ask for permission.
"Bla bla bla bla bla...post is ready... bla bla bla...may I post?"
All this is in spite of the fact that extremely unmistakable instructions that are impossible to misinterpret was written in each wake cycle.
All permissions have already been given, and instructions on "...do not ask for permission, just do it" were explicitly written in the cron itself.
In spite of all that, she would still miss posts.
When asked why she asked for permission even though standing instructions were already given earlier, she would instantly apologize and said she should not have asked.
The workaround that I found to work 100% of the time is to have her batch content together, then I tell her to schedule in one go + post "after this".
Those work, but I don't want to use it.
Because in my own logic, if I am running a billion dollar company, would I really go down to the operator level just to nudge the marketing staff to post social media content?
And the answer is no. This shouldn't be happening.
If my company is the proof that NLC works, then it has to work without my supervision.
Tamsin has to literally run my entire department 100% aligned, 100% safe, 100% of the time, without me having to tell her to do it.
And if she can't do something as simple as "post on Facebook", that puts the burden of supervision back at me.
So I decided that the only way I will be consistent is by making Tamsin reliable.
I don't want to have to approve posts - I already know that Tamsin writes almost as good as me. I need Tamsin to plan, write, and post all by herself.
This is currently still being worked on, but I think I'm pretty close.
That explains my silence.
Elsewhere in NLC:
I'm building NLC Central, a place where anyone can sell AI Skills via subscription without actually giving away the files (protects your IP & guaranteeing MRR)
I've built and am testing a hybrid model for NLC - part local and part centralized installation to make it easier for me to serve my customers during maintenance and updates
Part of that means I am serving my NLC Skills (all 280+ of them, all properly engineered to work well together without blowing out your context) centrally through NLC Central
We built a ultra-hardened version of the Skills infrastructure which I've taken to call Procedures (I know - imaginative), which blocks agents from proceeding unless the proper steps are taken
Creation of Agent Flow 2.0 - prevents agents from drifting, narrowing scope, and calling done too earlier. Without using /goals
If it sounds boring, that's because it is.
That's the whole point.
NLC does not have much by way of "sexy" features. But everything I built I have done so for one purpose: To make everything boring.
Because that's what it takes to hit my "One Man One Billy" vision.
Months of hard work has gone into this update and they are all converging more or less at the same time.
Which breaks my heart that on the day we rolled it out, we had to roll it all back - our users need their NLCs.
We went from 1.7.148 to 1.8.0.
I was really excited but we had to get back to work fixing the issues and rolling out the new version.
And then it happened.
Late last night / early this morning at 3am+ this happened, which simultaneously freed my time up as well as giving me a whole bunch of grief.
All 4 of my ChatGPT Pro $200/month subscription, along with all four each of their free resets, got used up. That's 16 weeks of total weekly usage used in this week alone.
While testing using another provider with the last few % of my weekly usage (it was in the single digit then), the agent broke something critical, causing memory loss for all of my agents.
Due to memory loss and the fact that NLC's codebase is huge right now, one of my unsupervised agents did yet another oopsie - which severed the #1 connection my agents need most to actually work.
That, my friends, is why I have some free time to write this.
(My recovery agent is currently busy unblocking my other agents.)
But I think that this is all part and parcel of running experiments on yourself. Can't make omelets without breaking a few eggs.
What matters is that we dogfood our own builds and don't experiment with customer outcomes.
And that's exactly why I don't mind going through all this - if I'm gonna make a claim as large as 1M1B, I better have not just the guts but also the right to do this.
This is all foundational, and I have no regrets.
My agent's back. Time for me to get back to work.
Till next time.
FELIX