We consider a service system where agents are invited on-demand. Customers arrive exogenously as a Poisson process and join a customer queue upon arrival if no agent is available. Agents decide to accept or decline invitations after some exponentially distributed random time, and join an agent queue upon invitation acceptance if no customer is waiting. A customer and an agent are matched in the order of customer arrival and agent invitation acceptance under the non-idling condition, and will leave the system simultaneously once matched (service times are irrelevant here). We consider a feedback-based adaptive agent invitation scheme, which controls the number of pending agent invitations, depending on the customer and/or agent queue lengths and their changes. The system process has two components-- `the difference between agent and customer queues' and `the number of pending invitations,' and is a countable continuous-time Markov chain. For the case when the customer arrival rate is constant, we establish fluid and diffusion limits, in the asymptotic regime where the customer arrival rate goes to infinity, while the agent response rate is fixed. We prove the process stability and fluid-scale limit interchange, which in particular imply that both customer and agent waiting times in steady-state vanish in the asymptotic limit. To do this we develop a novel (multi-scale) Lyapunov drift argument; it is required because the process has non-trivial behavior on the state space boundary. When the customer arrival rate is time-varying, we present a fluid limit for the processes in the same asymptotic regime. Simulation experiments are conducted to show good performance of the invitation scheme and accuracy of fluid limit approximations.