agentic workflows
/ a-JEN-tik WURK-floh /
An agentic workflow is a way of getting better results from an AI by breaking a task into steps the model works through over several turns, often with checking and revising along the way — instead of demanding one perfect answer in a single shot. The everyday analogy is the difference between blurting out a finished essay and actually writing one: outline, draft, reread, fix, repeat. Even a human genius writes badly in one pass; giving the model the same room to plan, act, and revise tends to lift quality a lot.
These workflows come in recognizable shapes. A model might plan first and then execute the plan step by step; it might draft something and then critique and improve its own work; it might split a job into sub-tasks and farm them out; or it might call tools, look at the results, and loop. Some workflows are tightly scripted by the developer (the steps are fixed, and the model just fills each one in), while more open ones let the model decide the steps as it goes — which is more flexible but also more fragile.
The trade-off is real and worth naming. More steps usually mean better quality but also more time, more money (each step is another model call), and more places for an error to slip in or compound. There is no single "correct" agentic workflow; building a good one is mostly engineering — deciding where to add a checking step, where to keep a human in the loop, and where a plain single call would have done just fine. The cleverness is in the structure around the model as much as in the model itself.
A "write code" workflow: (1) plan — the model lists the functions it will need; (2) draft — it writes each one; (3) test — it runs the code and reads the errors; (4) revise — it fixes what failed and runs again, looping steps 3–4 until the tests pass. The same model could produce far worse code if forced to emit it all in one shot with no chance to run and fix.
Plan → draft → test → revise: structure around the model, not just a bigger model.
More steps are not automatically better. Each extra model call costs time and money and adds a chance to fail; an over-engineered workflow can be slower, pricier, and no more accurate than a single well-prompted call. Add structure where it measurably helps, not for its own sake.