Building AI into a 3B workflow

An “AI step” is any step that calls a model. Pick the simplest shape that does the work, wire it the 3B way, and don’t hand-roll what the agent template already gives you.

Choose the shape

Simpler shapes mean less latency, cost, and breakage — but under-powering is just as real a failure: a task that has to gather context or act, crammed into one call, gives a worse answer, not a simpler one. The deciding question isn’t how much fits in the prompt; it’s whether you can name every input up front.

1. Single request — one call in, one result out, no tools. Use it when you can assemble the complete input ahead of time and the job is to transform it: classify, extract, summarize, generate, rewrite, answer a self-contained question. If a regex or lookup would do, don’t call a model at all.
2. Agent — one model with tools, looping until it’s done. Use it the moment a good answer depends on context the model has to choose to pull, or on doing something and reacting to the result: reading the files a diff touches, following a lead, searching, calling an API, then deciding what’s next. The tell: you can’t list the inputs up front, because they depend on what the model finds. Pre-fetching that context in code and pasting it into one call is the trap — it looks like it “puts everything in the prompt”, but it freezes what the model can consider to whatever you guessed and caps the quality of the answer. When the user asks for an “agent”, this is the shape they mean — and one pass is not an agent.

The agent step

Copy the agent template — never hand-roll the model loop, in-process or as a chain of AI steps. The template is the canonical implementation of what a correct agent guarantees: