Fixed workflow versus agent

2 min read

A fixed workflow is useful when the order of steps is known. It fails when the next step depends on what came back.

Same question: does QuickScreen meet both requirements? Both policies are written in advance.

Fixed workflow, k=1, budget 2

stopped

The model-call budget ran out. This is a technical stop, not proof that the collection has no answer.

Agent that continues after the observation, k=2

answered

Does not meet the joint requirement: recall of 90% is below 95%, and the time reduction of 40% is above 30%. Sources: P:1, C:1.

The same principle in dividing up jobs: one rule improves one input, and does not become a general rule.

Deterministic demo. No model is called here.

5Finish time

4Best for this input

4Lower bound

Worker 1 · load 5

A:2C:3

Worker 2 · load 2

B:2

The rule finishes at 5, but there is a split that finishes at 4. An improvement on one input doesn't prove the rule is always better.