On page 246 of the core, under the Agents heading, it says
An agent runs as a program and can use programs running on the same device as them.
Therefore even though it counts as its own entity it is still a program of a specific type. You can however have multiple different agent programs installed, but they cannot run simultaneously and requires a system reboot to switch since it counts as a persona. The section then goes on to state:
When an agent is running, it has its own persona (and icon). An agent is about as smart as a pilot program of the same rating (Pilot Programs, p. 269).
So to continue to explore that, the Pilot Programs section states that:
and
When faced with something novel or unexpected, or a complicated command, a Pilot program
must make a Device Rating x 2 Test
Therefore I would say that teamwork is possible with an agent, but you do so at a risk. If the target action has a high difficulty rating, the agent program might have to make a test just to attempt to help. With the right specific commands it can be prepared to help, but I imagine that would take some preset conditions to let it know when to pile on without trouble.