What we don't do
What you refuse to ship is architecture.
A short, honest list of the work we turn down — so you know what you'd be getting, and what you wouldn't.
The refusals
Production-grade, or not at all.
Production-grade, for an agent, isn't a better model or a cleverer prompt. It's the boring discipline software learned decades ago: you can measure whether it's working, you can roll it back, it knows the limits of what it's allowed to do, and someone owns it by name. So there are things we won't ship.
We won't ship an agent with no evals
If there's no way to tell whether it regressed, it isn't done. Measurement is part of the build, not an add-on.
We won't let unbounded autonomy touch what can't be undone
Irreversible steps get a human checkpoint and a rollback path, or they don't go live.
We won't sell you an observability platform
We use the instruments you already run, or open ones you keep. The judgment is the work, not the tooling — we're not a deployment arm for a vendor.
We won't automate what shouldn't be automated
Part of what you pay for is being told what AI should not touch in your workflow, and why.
We won't take on one-offs or staff augmentation
No experiments for their own sake, no feature factories. We work best when clarity and durability matter, on systems you intend to keep.
We won't deliver a strategy deck
You get running systems, a baseline, and a ranked plan in your own numbers. Nothing here is a slide for its own sake.
We won't widen past what we do well
One thing done properly beats five done shallowly. If we're not the right fit, we'll say so early.
If that fits how you think
Start with the audit, or the checklist.
If this matches how you think about systems you have to live with, we'll get along.