No-Nonsense Orchestration: Autonomous Agents on a SQLite Task Board

The Idea

I shipped an initial release of No-Nonsense Orchestration (58b51078): autonomous AI agent orchestration built on a SQLite task board. After a few iterations on orchestration design, I wanted the simplest possible coordination primitive—a shared table of tasks with states—rather than a message bus or a distributed queue. SQLite gives me transactional claims, visibility into what every worker is doing, and zero operational overhead.

Fixing the Migrations Layout

The first thing that broke was migrations. The code assumed a repo-relative migrations directory, which failed when the tool was embedded in another project. I fixed the lookup to work in both layouts (4a770513) and added the missing migrations directory with the initial schema (b43fb69a). It was a useful early lesson: if a tool can run from two locations, path assumptions will bite immediately.

GPU Exclusivity

The interesting rule came last: GPU exclusivity (a8f8501b). Local inference saturates the GPU, so running two model calls concurrently doesn’t speed anything up—it just makes both slower and risks out-of-memory failures. I encoded an exclusivity rule so only one agent can hold the GPU at a time while other agents continue on CPU-bound or non-inference work. That single constraint did more for throughput than any scheduling heuristic I might have written.

Why Simple Wins

The whole system is a table, a claim protocol, and a scheduling rule. That’s it. At this scale, a boring design is a feature—I can inspect the entire state of the machine with one SELECT, and there’s no broker to keep alive. I’ll add capability once the simple version hurts.