Agentic Workflows in Conversational AI

• Agentic workflows are structured, multi-stage processes that distribute decision-making tasks across specialized AI and human agents.
• They employ iterative design and embedding-based methods to clarify ambiguous user goals while ensuring adaptive, context-driven interactions.
• This paradigm enables cognitive offloading and prompt personalization, improving alignment between user intent and system capabilities.

Agentic workflows are structured, multi-stage processes in which autonomous agents—often leveraging LLMs—collaborate within complex decision-making pipelines. These workflows are distinguished by their ability to distribute cognitive and operational subtasks among specialized agents and human participants and to systematically orchestrate context, reasoning, and control across well-defined interaction stages. Advancements in agentic workflows address the core challenges of ambiguity, transience, and capability gaps in domains where human intentions are difficult to articulate and system affordances are not always transparent. In conversational human-AI interaction design, agentic workflows provide mechanisms for clarifying goals, suggesting contextually relevant actions, and supporting iterative, designer-in-the-loop engagement, thus enabling more robust, adaptive, and user-aligned interfaces.

1. Definition and Theoretical Foundation

Agentic workflows, as defined in the context of conversational human-AI interaction design, are formally structured, multi-stage exchanges that allocate both cognitive and operational tasks across humans (users and designers) and a variety of AI agent roles. The workflow is segmented into stages such as contextualization, goal formulation, and prompt articulation. At each segment, specialized agents (Contextual Persona agents, Proxy User agents, Goal Refinement agents) support the user, providing context-specific and role-driven interventions.

This division enables shared decision-making and fosters behaviors such as:

• Task Decoupling: Separating goal formulation from prompt articulation, enabling a transition from divergent, exploratory cognition to focused, actionable intent.
• Role-specialized Intervention: Assigning agent roles (e.g., context recommender, proxy user, goal refinement) to mediate ambiguities and operationalize transient interactions.

Mathematically, agentic workflows often rely on vector representations in embedding spaces for matching context to agent prompts. For example, the cosine similarity is used to select the most appropriate agent personas: $$\text{cosine\_similarity}(\mathbf{u}, \mathbf{v}) = \frac{\mathbf{u} \cdot \mathbf{v}}{\|\mathbf{u}\| \|\mathbf{v}\|}$$ where $\mathbf{u}$ and $\mathbf{v}$ are vector embeddings of the current conversational context and available persona data.

2. Addressing Ambiguity and Transience in Conversational HAI

Agentic workflows explicitly target two foundational challenges in conversational human-AI interaction (CHAI):

• Ambiguity: Users frequently articulate imprecise or evolving goals and lack comprehensive understanding of system capabilities (designated as the “capability gap”). The agentic workflow mitigates this via context-sensitive prompting, micro-hypothesis generation, and iterative goal refinement, systematically exposing latent or previously unarticulated user needs.
• Transience: Brief, task-focused conversational sessions limit the opportunities for iterative feedback and learning. In response, agentic workflows adopt incremental, annotated interventions that enable rapid hypothesis formation, prompt testing, and designer-in-the-loop experimentation, even within the constraints of ephemeral interaction.

3. Methodology and Iterative Design

The agentic workflow design leverages a research-through-design (RtD) methodology, comprising:

• Probe Development: Iterative creation and refinement of a chat-based AI web application powered by a lightweight LLM.
• Cycling and Evolution: Four documented cycles, each involving real users (n = 10), where feedback from usage logs, Likert-scale surveys, and qualitative input guides progressive adaptation of the workflow stages.
• Artifact Annotation: Portfolio-based tracking of design artifacts across versions (such as visual-context mechanisms, standalone goal formulation modules, and a designer interface for Proxy User agent interaction).
• Thematic User Analysis: Deep analysis of user experiences with focus on transitions between divergent/convergent thinking, the effect of agent-generated recommendations, and context overload versus informational sparseness.

4. Empirical Findings and Design Insights

Through qualitative and quantitative analysis, the following findings were documented:

Agentic Workflow Component	Empirical Outcome	Contextual Implication
Decoupling goal formulation	Facilitates transition to focused prompt articulation	Enhances alignment between user intent and system affordances
Specialized AI agents	Reduces ambiguity via context- and persona-specific prompts	Enables meaningful intervention even with initially unclear user goals
Iterative “micro-hypothesis”	Allows users/agents to propose and test hypotheses rapidly	Mitigates drawbacks of transient and brief conversational interactions
Overly rigid automation	Occasionally introduces irrelevant cues, misleads users	Necessitates correct calibration of contextual input for effective disambiguation

An iterative architecture—where contextualization, goal discovery, and prompt recommendation are developed as semi-autonomous agent roles—was found to enhance both user articulation and system reliability. A persistent theme is that the transition from undifferentiated conversational flows to staged, agent-mediated interactions supports better alignment between user needs and AI capabilities.

5. Limitations, Design Trade-offs, and Expansion Potential

While promising, the agentic workflow paradigm faces key limitations:

• Context Overload: Excessive contextual recommendations can overwhelm users or introduce misleading signals when automated context inference is imprecise.
• Sub-optimal Rigidity: Overly rigid or automated agent suggestions risk the propagation of errors in interpretation, stressing the critical importance of calibrated context and designer intervention.
• Scalability Constraints: The balance between automation and human-in-the-loop correction demands careful attention to avoid both overfitting to sampled feedback and underutilization of user-driven iteration.

Expansion opportunities include:

• Adapting agentic workflows for broader design domains, such as creative exploration or complex task automation, by porting the multi-agent, staged-interaction pattern to new verticals.
• Leveraging conversational AI probes as “needfinding machines,” systematically surfacing user requirements—even latent ones—in domains where explicit need articulation is rare or costly.

6. Systemic Implications and Practical Applications

Implementing agentic workflows in CHAI systems supports:

• Cognitive Offloading: By algorithmically partitioning planning and articulation, the system reduces user cognitive/motor demands, enhancing efficiency.
• Designer-in-the-Loop Adaptivity: Proxy User agents, constructed from empirical usage data, enable designers to simulate and pre-test user interactions, fostering iterative, user-centered design improvement even in sparse or ephemeral engagement settings.
• Prompt Personalization: Specialized role agents can employ embedding-based matching (e.g., via cosine similarity) to personalize recommendations and reduce the search space for users with ambiguous goals.
• Iterative Artifact Development: Annotated design artifacts afford a mechanism for tracking, grounding, and improving workflow interventions through successive cycles.

These features systematically ameliorate both ambiguity and transience, resulting in conversational AI workflows that are demonstrably more robust, interpretable, and adaptable compared to monolithic or naive prompt-driven systems.

7. Conclusion and Outlook

Agentic workflows, as instantiated in CHAI design, represent a rigorous, modular paradigm for structuring complex human-AI interactions. The design and empirical validation of such workflows affirm that:

• Segmentation into context, goal, and prompt stages, mediated by specialized AI agents, enables greater alignment with evolving user intent and system capability.
• Iterative, annotated, and designer-in-the-loop development cycles produce workflows that can adapt over time—coping with the challenges of ambiguity and transience endemic to real-world, task-oriented AI applications.
• By providing both practical design patterns and analytic frameworks (e.g., embedding-based agent selection), this approach opens scalable pathways for extending agentic workflows to broader classes of human-AI collaboration and complex system design.

The agentic workflow paradigm thus establishes a systematic methodology for balancing automation with user and designer control, setting foundational directions for both applied design practice and future research in conversational and collaborative AI systems (Caetano et al., 29 Jan 2025).