The present disclosure is directed to a mechanism for optimized application performance in SD-WAN, and includes the steps of receiving initial traffic packets at a first site for transmission to a second site; determining whether a direct tunnel is established between the first site and the second site based on a state of the second site, the state comprising an active state indicating that a direct tunnel is established between the first and second sites or an inactive state indicating that the direct tunnel is not established between the first and second sites; and in response to determining that the direct tunnel is not established, determining that the initial packets satisfy a configured trigger; forwarding the initial packets to the second site via a backup path; establishing the direct tunnel between the first and second sites; and forwarding subsequent traffic packets to the second site via the established direct tunnel.