The advent of ChatGPT has engendered a paradigmatic shift in the dynamics of task-oriented dialogue systems, eliciting a reevaluation of long-standing approaches to conversational agent architecture, design, and performance metrics. Prior to ChatGPT, task-oriented dialogue agents predominantly employed modular architectures, often rooted in classical intent-based and slot-filling techniques. These systems were commonly anchored by discrete components for natural language understanding (NLU), dialogue state tracking (DST), and natural language generation (NLG), each fine-tuned for specific and narrowly defined objectives. However, such architectures often demonstrated brittleness and exhibited a lack of generalization capabilities, making them susceptible to semantic noise and syntactic perturbations.

Conversely, with the proliferation of large-scale transformer-based models such as ChatGPT, there has been a migration towards end-to-end trainable architectures, which offer more fluid, context-aware, and semantically rich interactions. Utilizing mechanisms like attention and memory, these models implicitly encode a more nuanced understanding of dialogue semantics without requiring explicit intent or entity recognition modules. Furthermore, they can generate text conditioned on extended conversational history, thereby augmenting the coherency and contextuality of dialogic exchanges.

Moreover, the pre-ChatGPT era was characterized by a certain proclivity towards handcrafted or rule-based dialogue management strategies. In stark contrast, ChatGPT and its kin leverage data-driven methodologies, fueled by gargantuan corpora that encapsulate a plethora of linguistic nuances and colloquialisms. This has engendered a democratization of sophisticated conversational capabilities, circumventing the need for labor-intensive feature engineering or domain-specific expert knowledge.

In terms of user experience, ChatGPT and similar models have broadened the ontological horizons of task-oriented dialogue. Earlier systems were highly specialized, excelling in constrained, transactional tasks but faltering when faced with more exploratory, open-ended dialogues. ChatGPT's expansive knowledge base and conversational dexterity enable a synthesis of transactional efficiency and contextual profundity, allowing for a more multimodal interaction schema that blends task-oriented and chit-chat dialogue strategies.

However, this shift is not without its attendant challenges, including ethical considerations surrounding data provenance and the potential for model-generated disinformation. Additionally, there is the computational overhead and resource-intensive nature of training and deploying these large-scale models, a factor that may engender a form of algorithmic elitism, marginalizing those without the requisite computational wherewithal.

In summation, the introduction of ChatGPT has engendered a seismic shift in the landscape of task-oriented dialogue systems, influencing both architectural paradigms and experiential dynamics. This transition necessitates a comprehensive reevaluation of established norms, algorithms, and evaluation criteria in light of the capabilities and complexities introduced by these advanced dialogue systems.