Explorations / Agent systems
The model is the cheap part
Building a coding agent from first principles — first in Bun, now in Rust — because re-deriving a thing is the only way I really understand it.
What it taught me
Still buildingYou learn an agent’s design by trying to rebuild it from scratch.
Still open
If the model is the commodity, how much of an agent’s trustworthiness is just the harness around it?
Why rebuild it
The only way I understand a thing is to build it
I’ve been building a coding agent from first principles (first in Bun, now in Rust) not because the world needs another one, but because building a thing is the only way I understand it. You can read about an architecture for an afternoon and retain nothing. Re-derive it yourself and you meet every decision the real ones made, including the ones you’d never have thought to make.
You can read an architecture in an afternoon and keep nothing. Rebuild it and you meet every decision its authors had to make.
Where the work actually is
The model is the cheap part
The biggest thing it taught me: the model is the cheap part. The intelligence is nearly a commodity now; you call it. The expensive, careful work is everything around it — the harness that turns a thing which talks into a thing you can trust to act:
- Tool loop
- The call-observe-decide cycle that turns a model from something that talks into something that does.
- Retries & recovery
- What happens when a step fails, half-finishes, or returns garbage — most of the real code lives here.
- Sandbox
- The blast radius: what the agent can reach, and everything it must not.
- Permission model
- What it may do on its own, and what has to wait for a human yes.
- The deferral line
- The exact point where autonomy ends and it stops to ask — drawn on purpose, not where it happened to break.
Same model, different harness → capable vs. trustworthy
Fig. 01 — The model is the cheap core; everything that earns trust wraps around it
cheapthe human lineA capable agent and a trustworthy one are mostly the same model wrapped in very different harnesses.
What it taught me
When intelligence is a commodity, judgment is the product
That’s the same conviction my product work keeps landing on, in a new register. The impressive demo is easy. The system you’d actually let act on your behalf is a long accumulation of unglamorous judgment about when not to trust it.
When the intelligence is a phone call away, the judgment about when not to trust it is the whole product.
It’s the demo-is-not-the-product lesson, made structural. The model opened a door that used to be the hard part; everything that decides whether walking through it is safe now lives in the harness.