Predictable jobs suit fixed paths

An original example is processing invoices: extraction, validation, and review can follow a fixed path. Investigating an unfamiliar software issue may require changing direction after reading logs.

The second job needs more flexibility, but also stronger stopping rules.

Autonomy adds cost and new failure modes

More autonomy creates more ways to spend time or make an incorrect change. Compare a simple workflow with an agent on the same tasks before adding complexity.

Where the choice sits

A workflow is a program. You named the steps, the order, and what happens on a failure.

A model may write the text inside a step, but it does not pick the next step. That makes the path reviewable.

A reader can point at the box that sent the email.

An agent is a loop. The model sees a goal and a list of tools, then chooses a call, reads the result, and chooses again until it stops.

The path is different for two similar inputs. That flexibility is the feature and the risk.

A tool that was safe inside a fixed step can be called at the wrong time, or twice, because the model thought the first result was incomplete.

How to pick without a slogan

If a competent person would follow the same checklist every time, write the checklist. Put the model inside the steps that need language: classify, extract, draft.

Keep the branch conditions in code. You will spend less, and a failure will have a line number.

If the next action really depends on something you cannot enumerate, an agent may be the smaller design. Give it the few tools that match the goal, a budget on calls, and a stop when it repeats itself.

Do not start from a framework demo that connects every API you own. The demo is showing that calls are possible, not that the choices are good.

  • Write down who chooses the next step: code, or the model.
  • List the tools available on that path.
  • Cap the number of steps and the spend.
  • Log the chosen steps so a bad run can be replayed.

What to record before you rely on it

For each job, record whether a fixed checklist already exists. If it does, the default is a workflow with a model inside a step.

If it does not, write the goal, the allowed tools, and the stop condition before you connect credentials.

Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.

Do not relax the check until the model happens to pass. Name a person who owns that failure.

A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.

Common questions

Is a model that calls one tool an agent?

Not by itself. If your code always calls that tool at that point, it is a workflow.

It is an agent when the model is the thing that decides whether to call the tool.

Are agents more capable than workflows?

They are more flexible, which includes more ways to be wrong. Capability on a demo is not a reason to hand over a path you could have written down.

Sources & further reading

  1. Anthropic: Building effective agents ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories