How dumb could a software agent be and still be considered an intelligent agent, presuming that it can communicate with and take advantage of the services of other, more intelligent software agents?
This still begs the question of how we define or measure the intelligence of a specific software agent. Do we mean the raw, native intelligence contained wholly within that agent, or the effective intelligence of that agent as seen from outside of that agent and with no knowledge as to how the agent accomplishes its acts of intelligence?
We can speak of the degree to which a specific agent leverages the intelligence of other agents. Whether we can truly measure and quantify this leverage is another matter entirely.
In humans we see the effect that each of us can take advantage of the knowledge (and hence to some degree the intelligence) of others. Still, we also speak of the intelligence of the individual.
Maybe a difference is that with software agents, they are much more likely to be highly interconnected at a very intimate level, compared to normal humans, so that agents would typically operate as part of a multi-mind at a deeper level rather than as individuals loosely operating in social groups as humans do. Or, maybe it is a spectrum and we might have reasons for choosing to design or constrain groups of agents to work with varying degrees of interconnectivity, dependence, and autonomy.
So, maybe the answer to the question is that each agent can be extremely dumb or at least simple-minded, provided that it is interconnected with other agents into a sufficiently interconnected multi-mind.
But even that answer begs the question, leading us to ponder what the minimal degree of interconnectivity is that can sustain intelligence.