SQLite Is All You Need for Durable Workflows, And They Proved It

Hey everyone, welcome to Builder's Briefing for May thirty-first, twenty twenty-six. I'm Alex.

And I'm Sam. Good one today — the throughline is basically radical simplification, which, honestly, is my favorite kind of day.

Absolutely. We've got SQLite replacing entire workflow engines, a new mixture-of-experts model that only lights up one-eighth of its parameters, TypeScript compiling straight to native binaries, and a billion-dollar acquisition. Let's get into it.

Alright, the big story. Obelisk published this detailed breakdown of using SQLite — yes, plain old SQLite — as the backing store for durable workflow execution. Not a toy demo. They're talking production architecture.

Okay, so when they say durable workflows, they mean like the stuff you'd normally reach for Temporal or Celery plus Redis for? Multi-step orchestration, retries, state machines?

Exactly. Their argument is that SQLite's WAL mode, its deterministic replay, and the single-writer model give you all the primitives you need. And for workloads under about a hundred thousand concurrent executions, it's not just sufficient — it's faster and dramatically simpler to operate.

That's interesting because if you think about what most teams are actually building right now — AI agent pipelines, background job systems — almost nobody is at a hundred thousand concurrent executions. You're bound to a single writer, sure, but that constraint literally doesn't matter until you're way past product-market fit.

Right. And this reinforces a pattern we've been watching for like eighteen months now. The boring infrastructure stack — SQLite, single-binary deploys, embedded databases — is eating the microservices world from below. If you're reaching for Kubernetes and distributed queues before you have ten paying customers, this is your wake-up call.

Ship the simplest thing. Scale when you actually need to. I love it. Link in the briefing if you want the full breakdown.

Okay, shifting to AI and models. Liquid AI dropped a new mixture-of-experts model — eight billion parameters total, but it only activates one billion at inference time. Trained on thirty-eight trillion tokens.

Thirty-eight trillion! That's a massive corpus. And the one-B active parameter thing is huge for edge deployment and cost-constrained environments. You're getting near-frontier quality at a fraction of the compute.

Yeah, if you're deploying on-device or watching your inference bill closely, benchmark this against Mistral and Phi for your specific use case. It could be a game-changer.

What about that MCP piece? Quandri's team is basically asking if Model Context Protocol is already dead. That feels spicy.

It is spicy. Their argument is that adoption is stalling because it's too complex and doesn't clearly beat plain function calling. But I'd call this the contrarian take, not the consensus. MCP may still find its niche in multi-agent scenarios, so don't rip it out of your stack just yet.

Right, and what's wild is we also got Harness today — an open-source framework that automatically designs agent teams and generates their skills. You describe a domain and it wires up the specialized agents for you. That's exactly the kind of multi-agent scenario where MCP could still matter.

Good connection. And one more fun one — a startup called Shift is offering to clean your home for free to collect real-world training data for robotic manipulation.

Free cleaning! The data-as-a-service flywheel is clever, but yeah, obvious questions about what else gets captured when you've got robots and sensors roaming around someone's house.

Alright, dev tools. This one made me sit up. Perry compiles TypeScript directly to native executables — runs it through SWC for parsing, then emits native binaries via LLVM. No Node, no Bun, no Deno at runtime.

Wait, so you could ship TypeScript CLIs as single static binaries? That's kind of the dream for a lot of teams. Early stage?

Early stage, but the implications are obvious. And it fits right into today's theme — single-binary deploys, minimal dependency trees. Link in the briefing.

And speaking of simplification, the Zig build system got a major rework. More composable, less magical. If you've been eyeing Zig for systems work or as a C/C++ build replacement, this removes one of the biggest friction points.

Also worth flagging — Steve Yegge wrote a piece called 'The Last Technical Interview,' arguing that the traditional leetcode-style screening is finally dying because AI coding assistants make it meaningless. Whether you agree or not, his framing of what to test instead — system design, taste, collaboration with AI tools — is really worth internalizing if you're hiring right now.

That one's going to generate takes for weeks. But he's not wrong that the signal from a whiteboard fizzbuzz is basically zero now.

Okay, big acquisition news. Downdetector and Speedtest — two of the most recognizable utilities on the internet — sold to Accenture for one point two billion dollars.

One point two billion! For Ookla's properties. That's a massive validation of the crowd-sourced infrastructure monitoring model. Community-contributed data creating a moat that's worth a billion-dollar exit.

If you're building dev tools with network effects around community data, that's your exit comp right there.

Quick hits before we wrap. Snowboard Kids 2 has been fully decompiled — an absolute masterclass in N64 reverse engineering. There's a gorgeous interactive demo of the Voxel Space terrain rendering algorithm from the old Comanche games. And IXI's autofocusing lenses are nearing production — could eventually replace multifocal glasses, which is wild.

The autofocusing lenses! That's one of those things that sounds like science fiction until it just quietly ships one day. Love it.

So here's the takeaway for today. The throughline is radical simplification. SQLite replacing workflow engines, single-binary TypeScript, MoE models activating one-eighth of their parameters. The builders who ship fast are the ones resisting the urge to add infrastructure until the problem actually demands it.

Seriously — if you're building AI agents or orchestration systems right now, evaluate whether SQLite plus a single process gets you to your next thousand users before you reach for distributed anything. The teams winning right now have the smallest dependency trees.

That's the briefing for May thirty-first. All the links are in the show notes. We'll be back tomorrow — until then, keep it simple, ship it fast.

See you tomorrow, folks!