For the past 3 days I’ve gained a meaningful progress with the Overseer function of NLC.
I started the project for NLC as “NLC Stabilization Project”, but I soon realized that even something like this would take too lon because my app is getting large.
So, I decided to fold in Symphony and Sandcastle.
The rationale is this: Before I can even think about multi-agent orchestration, can the agents be trusted individually to create high quality work?
By high quality work, I mean:
Done means done. I don't have to discover from my own usage that its still broken or not working.
Does not introduce any other bugs
Does not require a 1 hour scoping conversation or repeated conversation about doing the thing
I want the ability to pass a vague objective to any agent, and trust that it will get it done to my liking.
That alone is already very sophisticated.
I am not talking about something simple like “build a website”
I am talking about complex tasks like creating an entirely new function in an already large codebase - without breaking anything in the process.
In other words, NLC agents cannot be a high quality toy but a production-grade enterprise solution for businesses with real stakes and minimal room for errors.
And it worked.
I was able to start parallel agents, hand them a vague objective, and from there I only needed two touchpoints:
Touchpoint #1: When they come back after doing their research and present a plan
Touchpoint #2: When they finish the job and I know “done means done” because I don’t see any issue with the thing they built/fixed
With this new system of persistent memory I was able to task my agents with very complicated tasks that requires multiple hours of work.
Screenshot 1: The CLI agent ran for 6 hours 28 minutes and 40 seconds.
This wasn’t the longest run.
I took this screenshot during pause (after I slept) at a 16.5 hour long run.
The reason why I am happy with that is because every time I paused and asked the AI questions about what it did, it demonstrated that it has followed my instructions precisely.
It never lost sight of the objective, the constraints, and the state of the matter.
But that is nowhere near enough.
There have been many technical debts accumulated over the months, and I constantly parked them for later for one urgent reason or another.
For my vision of having an AI corporate army, I must not be the bottleneck.
If I had to be the one to tell
So I pushed on, and while fixing the tail end of a major blocker to NLC’s stabilization project, I realized that the Overseer was the perfect way to dogfood this implementation.
(The blocker was a merge conflict issue for parallel workflows that consumed 8 days where I was only able to use just 1 agent at a time.)
I had a stroke of inspiration and decided to implement that,
Screenshot 2: This morning, I woke up to this.
These screenshots are from my WebUI.
Specifically, the sidebar of NLC’s WebUI which keeps track of all the agents and the incoming messages from them.
Almost all of these were started by the Overseer.
The instruction I gave it was to fulfill the contract with me comprised of 34 work items (complicated ones), and if it encountered a blocker or a bug in the orchestration process preventing the work from proceeding, it should create another agent to unblock that before continuing.
When I started this project, Overseer wasn’t able to do anything at all.
3 days ago, I had proof that a single agent is able to take a very vague objective of mine and turn it into reality and to my liking without much of my input.
2 days ago, after working through the night, the Overseer was only able to steer one agent to finish one of the 34 tasks.
Yesterday, it was able to complete 4 tasks of the 34 work items, but one new item has to be added to the list, making it 35.
In the past 12 hours, only 2 items were done because of two reasons:
Those two tasks were genuinely difficult
We found that the same guards that forces agents to behave is overzealous
So now as I am writing this, we are making changes to the Agent Flow framework to make it minimal-friction for agent tasks.
And then comes another round of dogfooding which is to make the Overseer drive them once again as I go to sleep - its 1am here.
Tomorrow morning I will get to see the results of it.
Stabilizing NLC and the Overseer will unlock a whole set of new possibilities for me and all NLC users.
Ultra long running objectives without using Ralph Loops and cron jobs.
No dumb repetitive work. Not a robot zombie slave. An actual leader who can drive projects forward autonomously.
A general, not an infantry.
And you guys get to be on the front row seat as I continue to develop this.
FELIX
P.S. Speaking of, I am hosting a workshop this Thursday: “Orchestration 101”
I will be sharing everything I’ve learned about Agentic Orchestration in the past 3 months.
If you’re interested, sign up here: [Archived link retired.]