Overview
Most consumer agents still behave like search boxes: they expect users to arrive with precise goals, correct vocabulary, and clearly stated constraints. Nidhi argues that this assumption ignores the articulation gap—the distance between a user's vague sense of what they want and a specification an agent can execute. Using shopping as the main example, she presents a collaborative loop with three phases. Discovery consolidates conversation history, personal context, references, hard constraints, inferred soft preferences, confidence levels, and real-time variables into a working state, then identifies the next question with the greatest information gain. Research maps user constraints to a domain ontology, selects an appropriate elicitation format, and performs comparisons, trade-off analysis, and summarization on the user's behalf. Response generation then adapts the presentation—such as bullets, comparison tables, or visual boards—to the user's task and desired action. Each phase requires dedicated evaluation, including fact retention, counterfactual sensitivity, question utility, turn efficiency, format accuracy, data fidelity, and actionability. The broader conclusion is that collaborative agents should accept incomplete intent, use shared visual language when words are insufficient, treat presentation as part of intelligence, and continuously improve the evaluators governing the full loop.
Sections
Core Concepts
Key terms underlying the collaborative-agent framework.
- Articulation gap: the distance between a user's vague preference or feeling and the explicit vocabulary and constraints required for conventional search or execution.
- Working state: the agent's evolving representation of the goal, session history, user context, hard constraints, inferred soft constraints, confidence levels, and relevant real-time variables.
- Intent gap: the set of unresolved variables or blockers that prevent the agent from producing a meaningful answer.
- Collaborative strategy: the agent's plan for choosing which missing preference or constraint to elicit next, ideally based on maximal information gain.
- Multimodal elicitation: selecting text, images, visual boards, comparisons, or interaction signals to uncover preferences in the format best suited to the constraint.
- Counterfactual sensitivity: an evaluation of whether changing a relevant part of the user's query changes the corresponding inferred constraint while unrelated constraints remain stable.
Broader Implications
Patterns and design implications synthesized from the framework.
- A collaborative agent is closer to a decision-support partner than a conversational search interface: it helps construct the request, performs the analytical work, and shapes the result for action.
- The optimal level of automation is likely funnel-dependent. Subjective, identity-linked exploration benefits from direct human participation, while objective comparison and transaction work is more suitable for delegation.
- Presentation selection and question selection are parallel reasoning problems: both require the agent to choose the format that maximizes useful information transfer.
- Micro-interactions such as clicks and hovers may improve preference models, but they should be treated as uncertain evidence rather than definitive expressions of intent.
- Merchant-provided ontology can improve domain accuracy, but a horizontal agent layer must preserve control over response composition to maintain a consistent cross-merchant user experience.
Recommended Actions
Concrete steps for teams building consumer-facing collaborative agents.
- Design the input experience and agent policy to accept incomplete goals, references, examples, and subjective language without forcing users to formulate search-ready queries.
- Create a typed working state that separates explicit facts, inferred preferences, confidence levels, unresolved variables, and time-sensitive data.
- Rank clarification questions by expected information gain and penalize unnecessary or repetitive questioning.
- Implement format-aware elicitation so objective constraints can use text while subjective preferences can use visual anchors or comparisons.
- Build and test mappings between user-facing concepts and catalog or domain ontology, incorporating domain-owner expertise where appropriate.
- Choose final response formats according to the decision task and ensure the information required for the next action is immediately visible.
- Establish phase-specific evaluators for state fidelity, calibration, question utility, turn efficiency, format accuracy, factual fidelity, and actionability, then evolve them alongside observed failures.