Free checklist
Most AI tasks need a workflow, not an agent
Five questions that pick the right pattern for an AI system, the signs that a workflow is no longer enough, and a checklist to run against a system you are building now.
Reliability compounds against autonomy
In a workflow, the path is fixed in code. In an agent, the model chooses its own path. Agents make impressive demonstrations, so teams often start there.
The arithmetic runs the other way. If each step succeeds nine times in ten, three fixed steps succeed about seven times in ten, and you can see which step failed. An agent choosing its own route through the same steps succeeds less often, and its failures are harder to trace.
For most business tasks the path is already known. Letting the model rediscover it on every run is paying for uncertainty you do not want.
Every workflow is built from these
Each is a graph of model calls with the edges fixed by code. Each call has one job, a defined input and output, and ideally a check.
Prompt chaining
A fixed sequence of steps, with checks between them. Extract the facts, draft the reply, check the style.
Routing
A first call sorts the input and sends it to the right handler. Easy cases can go to a cheaper model.
Parallel calls
Independent parts run at once, or the same task runs several times and the answers are compared.
Orchestrator and workers
A central call splits a task whose shape depends on the input. The gateway to agents.
Evaluator loop
One call produces, another checks against criteria, and it repeats up to a fixed limit.
Five questions, in order
- Does the task have a fixed sequence of steps? Chain it.
- Does the input fall into categories that need different handling? Route it.
- Can parts of the task be done independently? Run them in parallel.
- Does the shape of the task depend on the input in ways you cannot list? Orchestrate it, and constrain the orchestrator carefully.
- Is there a clear standard the output can be improved against? Add an evaluator loop, with a limit.
If none applies, a single model call with good context may be enough. That is a workflow of one step, and it often is.
The test for whether you have a workflow is simple: can you draw the graph before the input arrives?
The signs, and the middle path
A workflow fails when inputs keep doing things the graph did not anticipate. The signs are a prompt full of conditional branches, tools that overlap, and a growing list of special cases that each need a different sequence. The sign is never that an agent would be more interesting to build.
There is often a middle path. Keep the workflow for the ninety per cent of inputs it handles. Send the rest to an agent with a fixed budget and a person at the end. The workflow covers the volume, and the tail is small enough to watch.
Whichever you choose, keep the control flow in your own code, where it can be tested, and keep the model away from decisions a simple rule can make. See the four levels of AI.
Using the checklist
Is an agent ever the right answer?
Yes, when the shape of the task depends on the input in ways you cannot list in advance. Even then, give it a budget, limits on what it can do, and a person at the end.
Should we use an agent framework?
Only if you can still see and test the loop, the stop condition and the approval steps. A framework that hides them is convenient until something goes wrong.
Where should a rule be used instead of a model?
Anywhere the answer is the same every time: who must approve what, in what order, with what limits. A model is the wrong tool for a rule.
From pattern to production
ENGINEER chapter 7 covers each pattern in depth, the production principles that keep workflows honest, and a worked example. Or we can design the system with you.