Agents vs pipelines
If you can draw the steps, build the flowchart. The workflow patterns that should be your default — and the narrow case where an open-ended loop earns its keep.
The most valuable lesson about agents is when not to build one. A workflow — a fixed graph of model calls and tools — is cheaper, faster, testable, and cannot loop forever. An agent trades all of that for the ability to handle a path you could not write down in advance. That trade is worth it far less often than the hype suggests.
Start here
If you can draw the steps on a whiteboard, build the flowchart. Reach for an open-ended agent only when the number and kind of steps genuinely depend on what earlier steps return — and you cannot know them until you start.
Should this even be an agent?
Can you write the exact steps in advance?
Does each step depend on earlier results in ways you can't predict?
Can you cheaply check each step's output?
Is the number of steps unknown until you start?
Build a workflow — you can draw the steps.
A prompt chain, router, or parallel fan-out is cheaper, testable, and can't spiral. Add autonomy only where the flowchart breaks.
If failure is costly or hard to undo, add a human checkpoint no matter what the score says. This checklist is a heuristic — the honest default is the simplest thing that works.
Answer the questions and watch the recommendation move. Most real tasks land on "workflow", and that is the point.
The workflow patterns worth knowing
- Prompt chaining — feed one call's output into the next. Use when a task decomposes into fixed stages.
- Routing — a fast model (or a decision model) classifies the input and sends it down one of several fixed paths. Cheap, and it keeps each path's prompt focused.
- Parallelisation — run independent subtasks at once and combine the results. Latency falls; the pattern is still fixed.
- Orchestrator–workers — a lead call breaks a task into subtasks and dispatches them to workers. Fixed in structure, adaptive in content.
- Evaluator–optimiser — one call produces, another critiques, loop until good enough. Bounded by a fixed number of rounds.
What an agent actually adds
An agent adds the freedom to choose how many steps to take and what each one is — including steps the designer never anticipated. That freedom is exactly what you want for open-ended tasks (debug this unfamiliar failure, research this question from scratch) and exactly what you do not want for a task with a known procedure.
Careful
The honest default is the simplest thing that works, and for most products that is a workflow. When you do build an agent, bound it: a step and cost budget, validated tool schemas, and a human checkpoint before anything costly or irreversible. Autonomy is a dial, not a switch — turn it up only as far as the task requires.
Check yourself
Eduspheria wiki · Agentic AI, Foundations
0 / 4 answered
That is the foundation: a model, a loop, tools, and a context budget. The next chapter makes the loop smarter — planning a task before acting on it.