PaperlayerView exact revision on arXiv

Source-linked AI summary

AdaLens: Interactive Storyline for Monitoring and Steering Long-Running Agentic Data Analysis

Yangtian Liu, Yan Miao, Shuhan Liu, Yunfan Zhou, Dae Hyun Kim, Di Weng, Yingcai Wu

arXiv:2608.17834v1cs.HCcs.AI

TL;DR

Long-running agentic data analysis requires better ways to understand evolving analytical structure and intervene precisely during execution. AdaLens addresses this with a storyline-based monitoring and steering interface, and its case studies and user study indicate promising support for these activities.

  • Problem

    Analysts lack effective ways to trace evolving analytical structure and intervene precisely across concurrent, multi-step agentic analyses.

  • Method

    AdaLens combines a progressive storyline representation of plans, progress, findings, and data-column lineage with element-grounded steering and execution controls.

  • Results

    Case studies and a user study indicate that AdaLens supports analysts in monitoring evolving run state and steering ongoing, multi-step agentic data analyses.

  • Takeaways & Limitations

    The evaluation suggests that long-running agentic data analysis benefits from continuous monitoring and in-situ redirection beyond prompt–response exchanges.

  • Takeaways & Limitations

    Longitudinal, in-situ, and comparative studies of longer analytical runs are still needed to assess AdaLens in extended real-world use.

Abstract

from arXiv · show

Large language models are pushing data science toward increasingly autonomous and agentic workflows, with recent systems already supporting multi-step and long-running analyses. As these workflows become more autonomous, conventional interfaces no longer provide adequate support for two critical requirements: observability for understanding an agent's evolving reasoning and evidence, and steerability for redirecting low-value directions or deepening promising ones during execution. Existing interactive approaches improve process visibility and open intervention points, but they remain largely designed for discrete, turn-by-turn exchanges rather than the parallel branches and evolving decision structures of long-running agentic analysis. We study this need as interactive oversight in long-running agentic data analysis and present AdaLens, an interactive system for monitoring and steering ongoing runs. AdaLens combines a storyline-based representation that unifies analytical plans, execution progress, intermediate findings, and data-column involvement with steering interactions grounded in these analytical elements for directional guidance and execution control. We evaluate AdaLens through two case studies and a user study, examining how it supports analysts in monitoring and steering long-running agentic data analysis.

I. INTRODUCTION

As agentic data analysis becomes more autonomous, analysts need observability into evolving analytical processes and steerability over ongoing, branching runs. AdaLens addresses these requirements with a storyline-based representation that preserves analytical lineage and steering interactions grounded in analytical elements.

  • Motivation: Increasingly autonomous agentic systems shift analysts from step-by-step drivers to high-level overseers, making observability and steerability essential interface requirements.Observability concerns understanding explored directions, emerging insights, supporting evidence, and the run’s overall evolution.
  • Challenges: Long-running analyses create heterogeneous artifacts across interdependent steps, requiring legible representations that support both multi-granularity inspection and preserved analytical lineage.Analysts must follow branching directions while connecting findings to their data grounding as the process evolves.
  • Challenges: Steering ongoing analysis requires interventions at coupled levels, including redirecting high-level analytical directions and managing individual plan-thread lifecycles.Directional shifts can make active threads obsolete, while newly created directions may require corresponding execution changes.
  • AdaLens: AdaLens organizes analytical plans, summaries, and atomic insights along an evolving storyline timeline while threading data columns through it as persistent, traceable characters.This design supports multi-granularity inspection and preserves how findings remain grounded in data across analytical steps.
  • AdaLens: AdaLens provides storyline-grounded steering interactions that operate at both the intention level and the execution level.The system enables analysts to act on summaries and atomic insights to redirect analytical direction and manage execution.
  • Evaluation: AdaLens is evaluated through two case studies on real-world datasets and a user study involving representative analytical tasks.The evaluation examines the system’s support for monitoring and steering long-running agentic data analysis.

II. RELATED WORK · A. Empowering Data Analysis with LLMs · B. Visualizing Data Analysis Processes

Related work spans LLM-enabled data analysis and visual representations that help users inspect how analyses progress and results are reached. AdaLens extends this landscape by representing long-running agentic analysis as an evolving storyline linking temporal progression, multi-granularity inspection, and analytical lineage.

  • II. RELATED WORK: Related work examines LLM-enabled data analysis, visualization of analysis processes, and interaction with LLMs.
  • A. Empowering Data Analysis with LLMs: LLMs support more open-ended data analysis through broad domain knowledge and reasoning, while early tools generated visualizations from natural language and refined this mapping with scaffolding, examples, and multimodal interaction.Effective use still depends on prior analytical skill and careful inspection of intermediate results.
  • B. Visualizing Data Analysis Processes: Process visualizations help users trace and revisit how results were reached, with representations spanning multiple abstraction levels.
  • B. Visualizing Data Analysis Processes: At the data-transformation level, systems expose intermediate states through live profiling, animated pipeline explanations, inspectable query decompositions, streaming dataflow graphs, and Gantt-chart test visualizations.WaitGPT encodes data-shape changes with table glyphs, while ViseGPT uses Gantt charts for error localization.
  • B. Visualizing Data Analysis Processes: At the workflow level, flow graphs represent dependencies among data states and operations, while tree views structure branching exploration and context reuse.Examples include notebook dependency tracing, dataflow and task-flow graphs, Data Threads, hypothesis node-link diagrams, and branching narrative trees.
  • B. Visualizing Data Analysis Processes: Timeline and history views support reconstructing how results were reached by combining artifact and search interfaces or recording sensemaking provenance.
  • B. Visualizing Data Analysis Processes: AdaLens frames long-running agentic data analysis as an evolving narrative and uses a storyline visualization to jointly support temporal progression, multi-granularity inspection, and data-grounded analytical lineage.

C. Interacting with LLMs · III. INFORMING THE DESIGN · A. Problem Formulation

The paper motivates interactive oversight for long-running agentic data analysis, where analysts must understand evolving analytical structure and intervene during execution. AdaLens addresses this need through direct manipulation of visible analytical elements, informed by expert collaboration and a formulation centered on recurring analytical elements and their relationships.

  • C. Interacting with LLMs: Researchers have developed interaction techniques that give users more control over LLMs beyond standard chatbot interfaces.
  • C. Interacting with LLMs: The “gulf of envisioning” makes it difficult for users to anticipate how prompts become outputs, motivating post-hoc inspection and pre-output intervention.Related work includes provenance tracking, explanatory diagnostics, and studies of analyst verification behavior.
  • C. Interacting with LLMs: Other systems externalize plans, context, and semantic structures as manipulable objects, including inspectable plans, node-link graphs, abstraction hierarchies, and design-space maps.
  • C. Interacting with LLMs: Unlike turn-by-turn approaches with discrete checkpoints, AdaLens supports continuous intervention in agentic data analysis through direct manipulation of visible analytical elements.These interactions let analysts steer the evolving process asynchronously and in situ.
  • III. INFORMING THE DESIGN: AdaLens targets analysts using LLM assistants or agentic coding tools for multi-step investigations, where following intermediate progress and intervening when useful are important.The design was grounded through close collaboration with three researchers who actively use such tools in data analysis.
  • III. INFORMING THE DESIGN: Over approximately six months, weekly hour-long sessions with experts followed a design study methodology and examined how analysts inspect and intervene in ongoing agent-driven analyses.The discover phase used concrete analysis scenarios to identify information needed for effective oversight.
  • A. Problem Formulation: Agentic data analysis is formulated as repeated cycles of planning, execution, and synthesis in which an agent autonomously advances a high-level goal and dataset across an evolving run.The run persists beyond a single interaction and can pursue several candidate directions concurrently.
  • A. Problem Formulation: Interactive oversight requires analysts to understand an evolving run and intervene as it unfolds, using analytical elements connected through analytical lineage.The elements include steps, plans and plan threads, atomic insights, summaries, and data columns; lineage links plans to threads, insights, summaries, and motivating findings.

B. Challenges in Current Practice

Current practice presents three recurring challenges in long-running agentic data analysis: inspecting accumulating artifacts, tracing evolving analytical structure, and intervening effectively as runs unfold. Together, PC1 and PC2 characterize observability difficulties, while PC3 characterizes steerability difficulties.

  • Challenge overview: PC1 and PC2 characterize the observability challenge, whereas PC3 characterizes the steerability challenge.These are the three recurring practice challenges identified through ongoing collaboration with experts.
  • Observability: Analysts struggle to inspect accumulating analytical artifacts across run-level steps, analytical plans, findings, and supporting evidence.Current tools often do not organize artifacts for structured inspection, forcing analysts to locate and review relevant information across extensive intermediate outputs.
  • Observability: Analysts struggle to examine specific findings without losing sight of the run’s overall state.Intermediate conclusions can be difficult to inspect and follow up on in context, while scripts and intermediate artifacts may remain temporary and difficult to manage or inspect.
  • Observability: Analysts struggle to trace evolving analytical structures spanning concurrent directions and successive steps.Relationships among intermediate artifacts and their grounding in the underlying data remain implicit, making the run’s analytical structure difficult to reconstruct.
  • Steerability: Analysts struggle to intervene effectively as runs unfold, including pursuing promising directions, deprioritizing irrelevant ones, and adjusting unproductive plans.They largely rely on natural-language instructions, translating judgments and relevant analytical context into actionable requests.

C. Design Goals · IV. ADALENS · A. System Framework

AdaLens defines four goals for interactive oversight and implements them through an orchestrator–worker framework that preserves observability and steerability during autonomous, multi-step analysis. Its design spans multi-granular analytical representations, data-grounded lineage, contextual steering, and lifecycle control of parallel plan threads.

  • C. Design Goals: C. Design Goals: Four goals target observability and intervention at the levels of analytical direction and ongoing execution.G1 and G2 address observability, while G3 and G4 address intervention.
  • C. Design Goals: C. Design Goals: G1 represents active plan threads, step progress, summaries, atomic insights, supporting evidence, and involved data columns across granularities.The representation supports movement between overall progress monitoring and close inspection of specific findings.
  • C. Design Goals: C. Design Goals: G2 exposes lineage among plans, summaries, and atomic insights through shared data columns, enabling analysts to trace process evolution and shared grounding.Shared columns provide a basis for tracing divergence, convergence, and relationships within and across analytical steps.
  • C. Design Goals: C. Design Goals: G3 lets analysts steer analytical direction by acting on visible summaries, atomic insights, and data columns to prioritize, suppress, or deepen investigations.Steering is expressed through visible elements of the ongoing run, reducing the need to reconstruct accumulated context.
  • C. Design Goals: C. Design Goals: G4 supports creating, launching, pausing, modifying, and terminating plan threads while the overall analysis continues.These controls coordinate computational effort with changing thread value, emerging findings, and domain knowledge.
  • IV. ADALENS: IV. ADALENS: AdaLens instantiates autonomous, multi-step data analysis while preserving analyst observability and steerability throughout the run.The system framework section introduces coordinated views for understanding the process, interactions for steering and control, and implementation details.
  • A. System Framework: A. System Framework: AdaLens uses an orchestrator–worker architecture in which the orchestrator manages high-level reasoning and multiple workers execute specific analytical plans.Each worker contains an analyzer for iterative ReAct-style reasoning, code execution, interpretation, and reflection, followed by a summarizer that consolidates the process and outcomes.
  • A. System Framework: A. System Framework: The interface preserves plans, summaries, atomic insights, and supporting evidence for inspection while capturing analyst interventions in real time.This framework complements autonomous execution with ongoing observability and steerability.

B. Understanding the Ongoing Analytical Process · 1) Storyline View:

AdaLens supports observability of ongoing agentic analysis through coordinated storyline, chat, and inspector views, with the storyline view providing a unified temporal representation of analytical structure, progress, findings, and evidence. Its visual encoding and inspection interactions help analysts trace plan execution, column involvement, report details, and evolving analytical lineage across sequential and concurrent threads.

  • B. Understanding the Ongoing Analytical Process: AdaLens coordinates storyline, chat, and inspector views to reveal analytical structure and progress, accumulated findings, and supporting evidence.These views are designed to maintain observability of the ongoing analytical process for analysts.
  • B. Understanding the Ongoing Analytical Process: The storyline view treats long-running agentic data analysis as an evolving narrative and integrates diverse analytical elements into a unified temporal representation.It serves as AdaLens’s primary interface for elucidating dynamic analytical progression.
  • 1) Storyline View:: Plan threads appear as analytical cards showing descriptions and execution statuses, then become summary cards displaying completed-thread findings.The listed execution statuses include pending, analyzing, summarizing, completed, terminated, and failed; the supplied passage truncates the summary-card description.
  • 1) Storyline View:: Column instances join analytical events during execution, connect summaries and insights to involved columns, and merge, continue, or exit as steps progress.This metaphor lets analysts follow how columns dynamically participate while findings remain grounded in data.
  • 1) Storyline View:: Report anchors mark stage-level or final textual syntheses and link each marker to its corresponding detailed report in the chat view.They extend the storyline’s core analytical elements with compact access points to textual reports.
  • 1) Storyline View:: The layout arranges analytical steps left to right, connects elements through column-character trajectories, and alternates finding and converge regions.Horizontal placement encodes progression, while cards and atomic-insight glyphs indicate analytical-step and summary membership; trajectories indicate column involvement.
  • 1) Storyline View:: AdaLens incrementally applies ordering, alignment, and compaction to preserve trace continuity, reduce crossings and wiggles, and keep concurrent, nested storylines clear and compact.Unlike classical layouts, the setting may branch into concurrent analytical threads and contain nested finding structures.
  • 1) Storyline View:: Analysts can select plans, summaries, atomic insights, and report anchors to synchronize inspector and chat details, while highlighting, panning, and zooming support focused navigation.Selection activates the corresponding inspection state, and horizontal panning and zooming help users navigate large storylines.

2) Chat View: · 3) Inspector View: · C. Steering the Ongoing Analytical Process

AdaLens presents ongoing analysis through complementary chat and inspector views, then supports steering at both the analytical-intention and execution-control levels. The chat records workflow stages and outputs, while the inspector organizes intermediate insights and plan artifacts for detailed examination.

  • 2) Chat View:: The chat view represents an analysis run as a conversation list containing executed actions, agent responses, and generated reports.It also includes the initial analytical goal and subsequent user messages.
  • 2) Chat View:: System-generated chat entries expose distinct workflow stages, including orchestrator thinking, plan creation, dispatch, progress evaluation, finding synthesis, and report generation.Dispatched plans are presented with compact plan cards.
  • 3) Inspector View:: The inspector view moves analysts from process-level awareness to detailed examination of intermediate outputs through a coverage grid and detailed view.These two components provide complementary process-level and output-level inspection.
  • 3) Inspector View:: The coverage grid summarizes extracted atomic insights across insight types and dataset columns, using color intensity to encode coverage degree.This supports overview of how insights are distributed across analytical dimensions and data columns.
  • 3) Inspector View:: The detailed view adapts its inspection mode to the current selection, including plan-focused displays of descriptions, execution status, and live or recovered artifacts.The analysis stream can include generated code and textual outputs.
  • C. Steering the Ongoing Analytical Process: AdaLens enables direct interaction with meaningful analytical elements to steer analytical direction and control ongoing plan threads.These interactions support both intention-level steering and execution-level control, addressing goals G3 and G4.

1) Intention-Level Steering via Visual Elements: · 2) Execution-Level Control by Regulating Plan Threads:

AdaLens supports intention-level steering from storyline visual elements through FOCUS, IGNORE, and ELABORATE, alongside execution-level controls for regulating analytical plan threads. These interactions let analysts redirect promising or low-value directions, deepen specific findings, and manage plan-thread lifecycles.

  • 1) Intention-Level Steering via Visual Elements:: AdaLens provides FOCUS, IGNORE, and ELABORATE actions directly from storyline visual elements and passes the requests with selected-element information to the orchestrator’s next analytical step.FOCUS and IGNORE redirect subsequent planning, while ELABORATE supports narrowly scoped investigation of a selected finding.
  • 1) Intention-Level Steering via Visual Elements:: FOCUS prioritizes promising directions identified in findings or data columns for continued analytical investment.Analysts activate the FOCUS pen and select summary cards, atomic insight glyphs, or columns.
  • 1) Intention-Level Steering via Visual Elements:: IGNORE suppresses redundant, low-value, or misaligned directions before they consume additional time and tokens.Analysts use the same visual-selection interaction to suppress unhelpful findings, branches, or columns.
  • 1) Intention-Level Steering via Visual Elements:: ELABORATE deepens an important but underspecified finding by investigating its explanation, mechanism, and root causes without unnecessary unrelated branching.It targets summary cards or atomic insight glyphs, not columns, and uses a compact confirmation popover without the keyword chooser.
  • 2) Execution-Level Control by Regulating Plan Threads:: AdaLens provides execution-level controls on plan cards for creating, launching, pausing, modifying, and terminating analytical plan threads.These actions regulate the plan lifecycle and provide fine-grained control over analytical-task execution.
  • 2) Execution-Level Control by Regulating Plan Threads:: CREATE adds a user-authored plan thread, joining the current batch during dispatch or dispatching at the next analytical step otherwise.The analyst enters the new plan through a popover opened from a blank storyline area.
  • 2) Execution-Level Control by Regulating Plan Threads:: LAUNCH immediately starts a pending thread or resumes a paused thread from its preserved phase, while PAUSE suspends a running thread and preserves its current execution phase.Paused threads leave active execution until the analyst explicitly resumes them.
  • 2) Execution-Level Control by Regulating Plan Threads:: MODIFY revises a valuable plan’s formulation for later relaunch, whereas TERMINATE permanently removes an unhelpful thread from subsequent execution.Modification requires pausing a running thread first; termination is immediate for pending or paused threads and follows the current operation for active threads.

D. Implementation · V. CASE STUDY · A. Case 1: Goal-Oriented Analysis of Social Media Data

AdaLens is implemented as a model-agnostic web application that streams execution and steering updates in real time. Case 1 shows how analysts can monitor data involvement and findings while pausing, terminating, and extending parallel analyses toward a final synthesis.

  • D. Implementation: AdaLens uses a React-TS frontend and Flask-based Python backend with HTTP steering endpoints and Server-Sent Events for real-time events and artifacts.The architecture is client–server and model-agnostic by design.
  • V. CASE STUDY: The case studies examine monitoring and steering through storyline-based process progression, findings, data involvement, intention-level guidance, and execution-level control.The evaluation includes two case studies and expert interviews with external analysts experienced in LLM-based data analysis.
  • A. Case 1: Goal-Oriented Analysis of Social Media Data: Two candidate plans appeared in parallel, so ED paused both, terminated the categorical/boolean branch, and launched the numerical-variable branch.This let ED review candidate plans before allowing potentially unnecessary analysis to consume tokens.
  • A. Case 1: Goal-Oriented Analysis of Social Media Data: ED used a crawled Twitter dataset to investigate the ambiguous hot_degree metric and its relationships with other tweet, user, and engagement variables.ED is a data analyst with two years of social media data-analysis experience.
  • A. Case 1: Goal-Oriented Analysis of Social Media Data: The storyline showed hot_degree as the analytical focus, while the summary identified statuses_count, media_count, and followers_count as its strongest related variables.The three related columns also appeared with relatively large labels, indicating substantial involvement in the analysis.
  • A. Case 1: Goal-Oriented Analysis of Social Media Data: ED created a targeted follow-up plan to computationally model hot_degree from statuses_count, media_count, and followers_count while the main run continued.He used the storyline CREATE interaction rather than redirecting the ongoing run through a new chat prompt.
  • A. Case 1: Goal-Oriented Analysis of Social Media Data: After obtaining related variables and a plausible computational model, ED requested a final synthesis, and AdaLens generated a global report consolidating the findings.The report presented a coherent account of how hot_degree related to the other dataset columns.

B. Case 2: Exploratory Analysis of NBA Game Data

In an exploratory analysis of a high-dimensional NBA tracking dataset, EE used AdaLens to survey emerging findings, steer attention away from low-interest variables, deepen a promising matchup insight, and consolidate the results into a coherent game interpretation.

  • Stage 1: Starting with open-ended exploration: The storyline view let EE survey emerging findings, potential analytical directions, and involved data columns together as initial plans completed.Summary cards and atomic insights accumulated in the storyline and connected findings to data columns.
  • Stages 2–3: Steering exploration: EE used IGNORE to deprioritize two uninteresting ball-coordinate columns, which subsequently disappeared from the storyline view.AdaLens automatically sent a steering prompt to deprioritize the columns ball zone and ball z.
  • Stages 3–4: Steering exploration: EE used FOCUS on the “Bogut vs. James: Blocked-shot dominance” summary and ELABORATE on a temporal insight to pursue and explain their blocked-shot interactions.The insight stated that blocked-shot events between Andrew Bogut and LeBron James occurred substantially more often in the first period than in the third period.
  • Stage 5: Consolidating the exploratory findings: After the exploration, AdaLens generated a global summary report consolidating the surfaced findings into a coherent interpretation of the game.The report was requested from the chat view as a final synthesis of the key patterns discovered during the run.

C. Expert Interviews · VI. USER STUDY · A. Experiment Settings

Expert interviews found AdaLens valuable for visualizing analytical progress, preserving context, and directly steering runs, while identifying a need for clearer links across distant and inter-branch results. A task-based user study recruited 12 experienced students for tutorial and analysis tasks using AdaLens.

  • C. Expert Interviews: Both experts identified the storyline view as AdaLens’s most valuable component for showing process progression, step-level findings, and persistent column involvement.ED specifically valued size encoding for revealing which data were mainly analyzed.
  • C. Expert Interviews: IGNORE, FOCUS, and ELABORATE let experts continue from visible results without composing new instructions, making direct manipulation better than pure text control for ED.These interactions provided execution steering grounded in visible analytical results.
  • C. Expert Interviews: Both experts learned the interface quickly after a brief tutorial and requested no additional usage help during the case studies, suggesting a low learning barrier.The feedback concerns usability in real-world data analysis contexts.
  • C. Expert Interviews: AdaLens supported fine-grained plan-thread control, including terminating one thread without stopping the entire run, while helping ED maintain context.ED contrasted this support with tracking agent behavior through an evolving workflow.
  • C. Expert Interviews: Experts reported that distant related results can become difficult to connect and recommended explicit visual links and clearer inter-branch relationships.EE suggested representing whether results are correlated or causally linked to improve context management across longer runs.
  • VI. USER STUDY: The task-based user study evaluated AdaLens’s usability and collected participants’ experiences with the system.The study was organized around participants completing tasks with AdaLens.
  • A. Experiment Settings: 12 undergraduate and graduate students aged 20–26 participated; all had prior LLM and data analysis experience and represented STEM, social sciences, and humanities disciplines.Participants included 8 male and 4 female students, recruited through a university forum and compensated US$12 each.
  • A. Experiment Settings: The procedure began with a 15-minute tutorial and warm-up using a video game sales dataset, followed by 15-minute sequential analysis tasks on a student performance dataset.The warm-up familiarized participants with AdaLens’s interface and key functions before the task phase.

B. Findings · VII. DISCUSSION

AdaLens received strong usability, learnability, observability, and steerability feedback, while participants also proposed improvements for guidance and scalability. The discussion frames these results as evidence for continuous oversight, shared analytical representations, and further validation in longer, richer workflows.

  • B. Findings: The average SUS score was 87.08, exceeding the A+ threshold of 84.1, while usability and learnability scores were 87.76 and 84.38.All participants gave positive usability feedback, describing AdaLens as useful, intuitive, clear, and easy to use.
  • B. Findings: Participants suggested clearer hints for visual elements and interactions, more compact layouts, reduced clutter, and clearer ordering of important items.These suggestions identify opportunities to improve learnability and usability as the interface accommodates more analytical elements.
  • B. Findings: All participants (12/12) commented positively on AdaLens’s support for observing the analytical process through visible steps and intermediate findings.Many participants especially valued the storyline view for showing how analysis progressed and what insights it produced.
  • B. Findings: Many participants (9/12) expressed positive sentiments about direct steering interactions that enabled them to redirect ongoing analyses and refine goals on the fly.During free-form exploration, 9/12 participants independently used FOCUS or ELABORATE to initiate follow-up analysis.
  • B. Findings: All participants (12/12) felt AdaLens would better support long-running agentic data analysis workflows than a chat-only interface.Participants described AdaLens as supporting a different mode of work from repeated prompt–response exchanges.
  • VII. DISCUSSION: The evaluation suggests that long-running agentic data analysis benefits from interfaces beyond prompt–response exchanges, supporting continuous monitoring and in-situ redirection.Analysts tracked evolving run state, examined accumulated storyline context, and intervened during execution.
  • VII. DISCUSSION: AdaLens’s storyline unifies plans, summaries, atomic insights, and persistent data-column trajectories in a progressive temporal representation of analytical lineage.Plan cards become summary cards upon completion, while the layout incrementally extends as new analytical elements arrive and preserves rendered context.
  • VII. DISCUSSION: The case studies suggest that conversational framing combined with structured visual oversight can support directed follow-up and open-ended exploration.The discussion identifies potential applicability to automated report generation and iterative modelbuilding pipelines, while noting the need for longitudinal, in-situ, and comparative studies.

VIII. CONCLUSION

The paper presents AdaLens, an interactive system for monitoring and steering long-running agentic data analysis. It is built around a storyline-based representation and steering interactions grounded in analytical elements, and was examined through case studies and a user study.

  • AdaLens supports monitoring and steering long-running agentic data analysis through an interactive system.
  • Collaboration with three researchers experienced in agentic data analysis informed the problem formulation and identification of two key design challenges.
  • AdaLens combines a storyline-based representation with a progressive layout and steering interactions grounded in analytical elements.
  • The system was examined through two case studies and a user study.
Loading 2608.17834v1…