How we build agents: simple, verifiable, with clear tools.
What it is about
A practical guide drawn from many agent projects: which patterns work, when you actually need an agent, and when a fixed workflow is enough.
What caught our attention
The clear distinction between workflows (predefined steps) and agents (the model decides) - and the advice to start with the simplest solution and add complexity only where it measurably helps.
What we take from it
That is how we build: agents only where judgement is needed, everything else deterministic code. In our test operations with agents, two agents do the work - the rest is fixed logic with four human decision points.
