05.03.2026 13:14
π 0
π 0
π¬ 0
π 0
congrats!
we're using this "library" approach right now to build lifebuild.me and it's working pretty well. longer write-up coming.
early version is here if you want to peruse: github.com/sociotechnic...
I will say a *lot* more soon, but here's the gist:
- Software can be defined by documentation about the software (context)
- Software can be autonomously built to keep up with changes in the documentation
Working on building both the Library + the Factory
More here: sociotechnica.org/notebook/
the context is the software
the code is just a cache
every developer a 10x developer
I wonder if GitHub is ready for 100-1000x in volume
Godspeed!
Overnight on lifebuild.me I asked Codex to:
- Swap out my in-product agentic loop for pi.dev
- Refactor the entire UI, removing dead code
- Create custom project icons in-app w/ Nano Banana Pro
- Create a LifeBuild MCP server
π models β‘οΈ π ambition
Something people don't talk about enough when working with AI is just how much *fun* it is. It's not just about productivity!
Claude Code is fun.
Conductor is fun.
ClawdBot is fun.
Lean into the joy, folks!
First batch of alpha invites going out for lifebuild.me
Nervous, but excited! Feels so janky right now and there's a million things I want to fix, but really important to just get it out there.
bsky.app/profile/jess...
openclaw feels like a glimpse of the future:
- gateways to interact (right now, chat, but could be GUIs)
- tool-using LLM in a loop
- can bootstrap itself with new skills (discovered + new)
- can modify it's own code and manage its own config
why don't more apps work this way?
i mean i'd love to see more efforts like OCIF but failing that i think people could start cramming their app's data into HTML, PDF, etc. wouldn't make interop seamless but it would be easier to read and adversarially edit or script around
Same energy: bsky.app/profile/jess...
so cool! I still have an AE-1 at my parents house. would love to revive it.
We're talking hundreds of simultaneous agents.
For that, you need context to maintain alignment. And that takes a Library.
I'm not saying we the best solution yet, but I am saying this is going to be a MAJOR problem that more people need to start thinking about.
So far, I've written about the Software Factory and the Software Library.
Factory has gotten all the play. No one knows what to do with Library (or why they even need one).
People saying things like "I run 3-4 agents at a time, life is good! we've arrived!" That's not the factory. Not even close.
What will you do with your one wild lΜΆiΜΆfΜΆeΜΆ LLM-enhanced senior eng sw dev skillz?
The easiest path to malleable software: put an agent inside your app that has access to the repo for the product and can implement fixes and push PRs in response to the user pointing out problems.
"Self-healing software"?
But you're kinda right: Factories are the ultimate nerd-snipe.
"You mean instead of (having an agent) writing software, I can build a factory that (has agents) writing software for me?"
aren't we all? π
Full initial writeup with the five-dimension framework:
sociotechnica.org/notebook/co...
PS: If you're building in this space, I want to compare notes.
And we're building an AI librarian to maintain it.
His name is Conan.
Conan watches context streams, decomposes into atomic notes, establishes links, assembles slugs on demand.
His human boss makes judgment calls. Conan handles volume.
We're calling this a Context Library. There are two related context:
- The Library: complete knowledge base covering past, present, future
- The Context Constellation: curated context assembled for a specific task
Give an agent the full context and the micro-decisions improve. Not because the model got smarterβbecause the context got richer.
This is a management problem, not a model problem.
What they don't get:
- WHERE: What's adjacent? What breaks if you touch it wrong?
- WHEN: What did we try before? Why did it fail?
- WHY: What objective does this serve?
WHERE, WHEN, and WHY are the missing bundle.
None of this gets written down. Almost none of it reaches the agent.
Agents typically receive:
- WHAT: specs, requirements
- HOW: codebase, process instructions
The real issue is we've been starving agents of context human devs absorb by osmosis:
- Hallway conversations about why the last approach failed
- The senior engineer who says "we tried that in 2024"
- Ambient awareness of what's adjacent, what's fragile