Andy Smith

Building scalable and self-improving infrastructure for AI agents

[Andy Smith @andysmithai]

For years I've built and run production infrastructure at big companies with NixOS. At night I write Lisp for fun. At some point I thought: why not combine the two?

That's how Reflection started. An infrastructure framework for AI-native companies, built on Zeno, a self-evolving agent orchestrator in Clojure Lisp.

The trouble with personal AI assistants is folding each result back into the team's shared context. Everyone puts out a huge stream of code and docs, and the usual ways of merging it (PRs, reviews, tests) start to fail. Most of the context stays inside personal agents that never sync.

It's a bit like a car factory that skipped the assembly line. Every worker takes a blank home at night and brings back a finished part. Nobody knows how they did it, nobody can judge the process, and you can't see the hidden defects from the result. No serious factory works that way. In software, it happens all the time.

A company is its own thing. One assembly line, with its own context and its own set of agents. Not just a group of people who each do a piece.

This idea needs proper tooling. Reflection and Zeno let you describe a whole company as code. The orchestrator can create agents on its own, from fixed logic or from what other agents produce. Each one runs in an isolated sandbox with role-scoped access to the shared context, and the roles change on the fly.

People work through chat: Zulip, Discourse, Buzz. They discuss ideas, assign tasks, and answer the agents' questions. The company's agents do the work.

An organization can evolve this way, but people set the rules of that evolution. And Lisp makes that elegant, which brings the fun back into development.

[Justin Kim @justinplaygame]

We built Figma for hardware.
https://t.co/PskInke3Kd

[Andy Smith @andysmithai]

@CittadhammoB @Modostech @FrameworkPuter @dhh Great! Unbelievable for an e-ink display!

[Python Hub @PythonHub]

Hyperresearch

Agent-driven research knowledge base. Agents collect, search, and synthesize web research into a persistent, searchable wiki.

https://t.co/EMaBTVgJt9

[Andy Smith @andysmithai]

@codewithkhalil Budget 💸

[Andy Smith @andysmithai]

@ThePeterMick It should also have Apple Pay to buy extra usage quickly

[Andy Smith @andysmithai]

@Malix_Labs @ronefroni @mightyiam @oppili @tangled_org @vaxryy Nixvana is the best word to describe this state of mind

[Andy Smith @andysmithai]

@josephfounder @mark_l_watson Another good starting point is SICP: https://t.co/z8lEls0MJf

[Andy Smith @andysmithai]

@MikeLeiterDev Custom setup based on Clojure LISP

[Andy Smith @andysmithai]

@Abhijit_091 It might be cheaper (and much more private) to use a local model to obfuscate the request, send it to an expensive model, and then deobfuscate the response using the same local model, doing all of this on the fly at the LLM gateway level

[Andy Smith @andysmithai]

@YichuanM The near future is UX that adapts itself to the user’s needs

[Andy Smith @andysmithai]

Given how many architectural pitfalls I spot every day and make my agents think about, I genuinely don’t understand how vibe coders can just ignore them.

At the same time, I’m cautious about relying on AI for things I’m not an expert in, especially when I don’t know how to reliably verify the quality of the result or the process that produced it.

[Andy Smith @andysmithai]

@oldstackjournal Zulip is trying to combine the best of both worlds

[Andy Smith @andysmithai]

@bot Trusting AI with money seems like

https://t.co/X92WbWNAjr

[Simplifying AI @simplifyinAI]

MICROSOFT JUST OPEN-SOURCED SELF-EVOLVING AGENT SKILLS.

it's called skillopt.

skills that improve themselves the same way you train an ai model.

no more guessing what changes will actually help your prompts.

the base model runs the task, then an optimizer evaluates the output and rewrites the instructions itself.

→ seperates successful paths from failures to find actual improvements
→ auto-rejects any edit that doesn't exceed your benchmark score
→ beats hand-crafted prompts and optimizers
→ zero model lock-in, the skill transfers to any model

100% free. open source.

[Andy Smith @andysmithai]

@rfgarcia It will happen naturally once most customers are agents

[Andy Smith @andysmithai]

@livingdevops It’s different, but for me it’s better because I don’t need another language to describe cloud infra

[Andy Smith @andysmithai]

Yeah, separating public, personal, and company knowledge is a serious problem. Anyone who tries to work in public runs into it. It’s way too easy for something private to leak out.

I don’t know what the general solution is, but I use a concept I call “workshops.”

A workshop is a separate environment, completely isolated from everything else. In my case, it’s a Mac Studio. The agents running there have no access to anything private, so everything they can access is public. That means they’re physically prevented from accidentally publishing something they shouldn’t.

[Andy Smith @andysmithai]

@cnunciato @livingdevops Yes, I’ve used Terraform Cloud for years

[Andy Smith @andysmithai]

A couple of days ago I decided that every day I would post one result from that day. Something I built or changed in the real world. Something I’m actually proud of today.

But almost immediately I realized that it’s pretty hard to show off something that looks like sh*t.

So today I can finally say that I started working on my visual identity and design system.

I’m not a designer, so I won’t say it turned out amazing, but I like it.

Minimalistic and nice.