The log/

For some builds, managed platforms are the right move. They let you ship faster and remove noise. For other builds, the core system should be close to the metal: model gateway, data store, queue, logs, evals, deployment.

We like open-source when it makes the product more understandable. When the system is central to the business, ownership is not a luxury. It is how you debug, adapt, and keep moving.

We start with the job: classify, write, search, reason, extract, call tools, or decide. Then we test the failure modes. Latency, price, context, privacy, deployment, and eval quality matter as much as raw intelligence.

A small local model can be perfect for one step. A frontier model can be worth it for a hard decision. Most real products need routing, not model worship.

A good agent does not pretend to be magic. It knows its tools, its budget, its confidence, and the moments where it should stop and ask.

The product work is in the rails: approvals, logs, replay, escalation, memory, rollback. That is where trust is built. Not in a prompt, in the system around it.