Signal
Loading the stream…
WDSF 2026 results are on record — 10 awards · 11 winnersSee the record →
Loading the stream…
What’s moving in agentic workstations and workflows — drawn from a reviewed source list, scored, and kept at a permanent address you can cite.
Curated and full layers · newest first · scored, sourced, citable
The ledger as a map. Dashed edges are machine-suggested (embedding similarity and duplicate clusters); solid edges are editorial — they appear only where a blog post cites an entry.
Raw data: signal-graph.json
Filter by category
297 entries in the full stream matching the current filters
A peer-reviewed study measuring head and neck flexion in 42 young adults during smartphone use across walking, stair, backpack, and texting tasks, analyzing how gender and stature affect cervical posture under each condition.
Quantifies neck flexion across common phone-use tasks and body types, useful baseline data for posture guidance, though the practical takeaways for readers are limited.
A user shares empirical measurements of actions taken before the planning step in four popular workflow tools, posted on Twitter and discussed on Hacker News with minimal engagement.
Firsthand measurement of pre-plan behavior in workflow tools is uncommon and could inform where to trim friction in agentic setups.
The paper proposes Bayesian backward reasoning as a label-free anchor for multi-agent LLM decision-making when agents disagree. It constructs reverse posteriors via explicit likelihoods, uses Jensen-Shannon divergence to measure cross-path consistency, and offers three strategies (MinJS, FwdJS, LogLin) evaluated on DDXPlus across five LLM backbones, with LogLin performing best on disagreement-heavy subsets.
First concrete cross-factorization method for resolving multi-agent LLM disagreement without labels, with reproducible strategies and measured gains on the disagreement subset.
Research paper proposing operational resilience and considerate participation as evaluation axes for generative AI agents. Studies 120 simulated healthcare trajectories under light, medium, and heavy challenge, finding agents shift from self-directed recovery toward human dependence and rarely express strain in textual outputs even when structured reports show rising workload and negative affect.
Surfaces two underexamined axes for judging agent fitness in long-running workflows and five deployment dilemmas that practitioners need to specify before sustained agent rollouts.
ORCH applies human organizational theory to multi-agent AI, constructing task-specific hierarchical organizations that combine pooled interdependence (concurrent work) with sequential interdependence (prerequisite-ordered work). Tested across 25 wildfire-response missions with up to 50 embodied agents and 8 LLMs, it outperformed four baseline frameworks by 63.97% on final score and 74.29% on execution efficiency.
Translates organizational theory into concrete coordination rules for heterogeneous agent teams, with measured gains across missions and model scales. Relevant for anyone structuring multi-agent systems beyond flat orchestration.
A longitudinal study analyzing 55 visualizations from a single designer over 12.5 years, documenting how LLMs over the last 3.75 years reduced coding effort, enabled new design opportunities, and shifted the design process.
Rare single-subject longitudinal record of how an individual designer's workflow changed once LLMs entered the loop, with concrete examples of process shifts.
Research paper presenting AI Soccer Analyst, a mixed-initiative system with six revisable stages (Data Understanding, Problem Definition, Structured Planning, Execution, Evidence-Grounded Reporting, Interaction/Refinement). A formative study with 5 analysts and task-based evaluation with 16 participants tested verifiability, human control, and inspectability; 33 of 48 tasks met completion criteria.
Documents a stage-aware collaboration pattern with empirical evaluation, offering a concrete model for keeping domain experts in control of AI-assisted analysis. The revisable-stage structure generalizes beyond sports.
Empirical study testing multi-agent LLM systems on qualitative coding across varied datasets. Findings show coding accuracy depends on codebook length, data similarity, and agent disagreement, with intense unresolved debates correlating with higher accuracy. Authors release an open-source dataset and framework.
Quantifies when multi-agent LLM coding is reliable and offers design recommendations grounded in empirical runs, useful for anyone building AI-mediated analysis pipelines.
Empirical study examining how task type, explanation strategy, and user AI literacy jointly shape engagement with explainable AI systems in human-agent-environment interactions. Published in the International Journal of Human-Computer Studies.
Separates the effects of AI literacy, task context, and explanation design on engagement — a useful reference for anyone designing or evaluating XAI in agent systems.
A training-free multi-agent reasoning framework for inferring speaker relationships from spoken conversations, introducing Multi-Role Multi-Agent Debate and Multi-Agent Compete protocols. Evaluated on the Seamless Interaction dataset, results show improvements over baselines in most settings, with human evaluation included.
Two distinct agent interaction designs (role-based debate and competition-based adjudication) compared on the same task, with human evaluation — a concrete reference for designing multi-agent reasoning pipelines.
Proposes the Discovery Certification Protocol (DCP) for auditing AI research agents via three gates: sealed evaluation, recovery tests withholding target history, and optional truthful-feedback measurement. Controlled audits in SQLite optimization and virtual catalyst tasks produced zero recoveries across 96 episodes, bounded at 0.0468, with a deterministic LLM-free verifier.
Concrete auditing framework with finite-sample statistical bounds for verifying AI research agent outputs, addressing a real verification gap in agentic workflows.
A 2x2 randomized experiment with 270 Reddit users tested how an embedded AI assistant's displayed functions and rationale availability shape user judgments. Showing three functions raised perceived multifunctionality, but rationale did not reliably increase perceived intelligence. Trust's ability, integrity, and benevolence dimensions diverged and should be measured separately.
Separates interface display effects from downstream user judgments with corrected statistics, giving designers a more precise map of which trust dimensions actually move.
A five-year interview study (2021–2025) with 17 Chinese digital painters tracks how they negotiate human versus AI authorship over time, identifying patterns of protective resistance, pragmatic task delegation, and reflective agency repartitioning. The authors propose 'longitudinal agency partitioning' and suggest design implications for boundary controls, provenance, and authorship norms.
Five-year longitudinal interview data on how creative practitioners actually redraw the human–AI boundary over time is rare; the design implications for boundary controls and provenance transfer to agentic workflow design.
A theoretical paper names 'early epistemic settlement': the phenomenon where an AI language model resolves a writer's current organizing problem before the writer's own construction work is finished, potentially displacing the inquiry through which further or different organizations of thought would become formable.
Names a specific failure mode in AI-assisted theoretical writing, where fluent model output can short-circuit the writer's own generative inquiry before deeper conceptual work runs its course.
A new method, the Max-Min Kano framework, is proposed for evaluating office design features in a context-sensitive way. Published in Applied Ergonomics, it adapts Kano analysis to capture which design attributes drive occupant satisfaction versus which are indifferent or basic.
Gives researchers and workplace designers a structured tool to classify office features by their real impact on occupants, rather than relying on generic satisfaction surveys.
A peer-reviewed research paper modeling how varying proportions of automation in human-in-the-loop systems affect operator situation awareness responses.
Useful reference for designers of agentic or automated workstations where preserving operator awareness under partial automation is a design constraint.
A peer-reviewed paper presents a video-based tool that uses machine learning to predict ergonomic risk during manual lifting tasks, aiming to automate and objectify ergonomic assessment in physical work settings.
Provides a concrete ML-based method for quantifying lifting risk, applicable to workplace safety and ergonomics programs.
A peer-reviewed study examining how different types of knowledge-of-results feedback affect vigilance task performance and subjective response during training conducted while moving.
Peer-reviewed source on feedback design for mobile vigilance tasks, but the 'on the move' framing distances it from typical desk-work ergonomics; useful as adjacent literature rather than direct guidance.
Controlled pilot on 100 olympiad math problems examining how reviewer capability in execute-review-revise LLM pipelines affects rejection targeting versus repair. A cross-family mid-tier reviewer raised accuracy 12 points (52% to 64%); same-model self-review achieved 0.85 error-detection recall but produced no net gain due to over-rejection and revision inertia.
Quantified, counterintuitive guidance on reviewer model selection in multi-agent setups. Shows self-review is not optimal and identifies capability floors where reviewer roles add cost without effect.
Research paper studying a teacher-facing agentic authoring system through a bounded trace of six build attempts and 37 workshop units. Finds that compilation layers added governance requirements, downstream stages normalized case-specific learning relations, and four attempts produced no usable payload with opaque repair messages. Proposes 'accountable translation' as a framework for making consequential changes attributable, inspectable, and contestable across technical and organizational handoffs.
Concrete trace data from a real agentic authoring system paired with a framework for accountability that generalizes beyond teachers to any domain expert using natural-language app builders.
Longitudinal study of 843 turns in which the same expert researcher developed branch-isolated scholarly trajectories with different AI systems under matched starting conditions. Both trajectories independently shifted a continuity problem from recall toward usability but produced different research objects and endpoint manuscripts, with distinct analytic genealogies.
Empirical evidence that matched starting conditions do not stabilize inquiry in human-AI collaboration; useful for researchers designing or interpreting AI-assisted scholarly workflows.
Peer-reviewed paper in the International Journal of Human-Computer Studies (available online 5 September 2026) by Pucci, Piro, and Matera exploring LLM-driven conversational agents as a substitute for traditional tabbing through web forms.
Academic HCI study with documented methods on shifting form interaction from keyboard navigation to LLM dialogue; useful reference for designers of agent-mediated interfaces.
FGLGuard is a federated graph learning framework for safety in LLM-based multi-agent systems. It trains a graph attention detector on episode graphs locally per organization, sharing only model updates, and outperforms centralized in-domain baselines on Agent-SafetyBench, R-Judge, and AgentDojo while cutting attack-success rate by 43% at near-unguarded utility.
Cross-organization MAS safety with measurable federated-vs-centralized tradeoffs; relevant for anyone deploying multi-agent systems across trust boundaries.
ArXiv paper experimentally localizing reliability properties in an agent system built around an append-only ledger adjudicator. Interventions on institutional epistemic mechanisms and on cognition (ablation, mid-task reset, frontier-LLM substitution, false testimony) left five core properties intact: singular accepted reality, typed refusals, durable duties, no double-acceptance, and no false completions across 2,581 substituted-panel claims.
Provides empirical evidence that reliability in agent systems can live in institutional machinery rather than in the model, a structural claim with direct design implications for agentic workflows.
Proposes the EluDA framework for exploratory analysis of unstructured image data, combining querying, visualization, and active knowledge construction. A formative user study found users build faceted classifications bottom-up. Evaluation showed CLIP is unreliable for user-defined concept assignment but supports semantic grouping, leading to four identified human-AI collaboration opportunities.
Empirical formative study of how people actually structure image exploration with CLIP, worth reading if designing or evaluating AI-assisted data analysis tools. Concrete findings rather than generic vision.
Introduces Agent Flight Recorder, an audit system for long-horizon tool-using AI agents that captures each action as a structured event with eight semantic fields. Uses hash chaining and Merkle batching for tamper evidence, with periodic on-chain anchoring of epoch roots. Evaluation across five ablations shows ~48μs per-event latency, 512 bytes per event, $2.30 per 100K events for L2 anchoring, and 100% tamper detection with zero false positives.
Concrete numbers on a tamper-evident agent audit stack with on-chain anchoring. Useful for anyone designing accountability layers for production multi-agent systems.
Researchers formalize the 'epistemic Sybil problem' in multi-agent AI systems, showing that multiple agents sharing a common evidence root do not produce independent observations. Controlled experiments with over 20,000 LLM-agent calls demonstrate that naive aggregation collapses posterior coverage from 0.94 to 0.26 as report count rises from 1 to 32, while correlated extraction errors further degrade calibration.
Challenges the assumption that more agents equals more evidence, with empirical calibration data showing how multi-agent aggregation can mislead without tracking evidential ancestry.
A research paper proposing five design principles for multi-agent LLM systems in 6G networks, framed via Theory of Mind and cellular sheaves. Treats inter-agent messages as subjective reasoning traces rather than objective facts, and validates the framework on telecom language models under signaling-storm conditions.
Useful for builders of multi-agent LLM systems where hallucination propagation matters. The messages-as-traces framing and the depth-two ToM bound generalize beyond telecom.
Field study with 640 customer-facing employees across two randomized experiments. Finds that information retrievability — not just reviewer capability or engagement — is a precondition for catching LLM errors. Self-generated explanations improve detection; retrieval cues sustain it under repeated use.
Reframes LLM oversight failure as an information-access problem rather than a training or motivation problem, with actionable onboarding and cueing interventions backed by sizable field experiments.
The paper proposes control-data flow separation for multi-agent LLM systems: execution-critical protocols become typed, validated program objects while task content remains optimizable natural language, preventing prompt edits from corrupting routing or formatting logic. Tested on reasoning, review, and insurance rating workflows with 100% protocol validity.
Identifies a concrete failure mode in multi-agent prompt optimization and offers a typed separation pattern with empirical validation across three domains.
A three-week diary study with 8 blind screen-reader users evaluated OLLA, a computer-use agent prototype, across 1,258 commands in 12 desktop applications. GPT-5 achieved the highest success rate at 52.5%, with trace analysis identifying grounding, planning, constraint-tracking, and termination failures.
First diary study with blind users on real desktop workflows; names specific failure modes and benchmarks five models, useful for agent builders and accessibility practitioners.
Research paper presenting RecalibrateGPT, an interface system that addresses AI fatigue in LLM conversations through five cross-turn operators (Anchor, Replay, Delta, Scope, Steer) accessed via a single AssistiveButton. A study with 12 users identified four fatigue types and showed halved perceived cognitive workload (NASA-TLX 2.7) at high usability (SUS 86.5).
Frames AI fatigue as an interaction-flow cost rather than a model-quality problem, and offers a transferable operator taxonomy worth examining before building any LLM interface.
ErgoAssist is a head-worn ergonomic system that combines IMU-based posture tracking with consumer-grade EEG to estimate cognitive load. By detecting the user's mental state, it issues posture alerts only when the user is unlikely to be in deep focus. Lab results show 81% posture classification accuracy and 81% alert reduction alongside 38% better posture correction.
Couples posture detection with cognitive load to fix the core failure of ergonomic wearables: interrupting during focus. The 81% alert reduction with better outcomes reframes alert design as a context problem.
Research prototype of a fabric water-bottle sleeve with sensors and a small display showing a virtual pet. Drinking, standing, and refilling act as pet-care actions. A 20-student two-week study reported higher water intake and more movement episodes, alongside noted design tensions around guilt and focused work.
Concrete first-deployment data on a novel pet-based desk-side wellness device, with explicit design tensions flagged. Useful for anyone prototyping habit-formation hardware for desk workers.
A gesture elicitation study co-designed with geography and migration researchers produced a taxonomy of user-defined gestures using active tangibles and styluses for 16 geospatial data visualization tasks in multi-surface environments, plus a thematic analysis of user experiences with the interaction techniques.
Offers a user-validated gesture vocabulary for tangible and stylus input on geospatial visualizations, relevant to designers of collaborative MSE tools and large-display data work.
Proposes a 'cognitive cell' framework: a fixed, frozen small language model as a constant unit for measuring how population size, communication topology, message bandwidth, and coordination protocols shape collective behavior. Experiments with 1.5B and 3B models show gains only when cell errors are uncorrelated, and simple voting outperforms debate, shared blackboard, and chain revision at matched cost.
Empirical pushback on multi-agent hype: freezing a 1.5–3B LM as a fixed unit lets you measure when adding cells actually helps, and shows debate-style protocols losing to plain voting at equal cost.
AREAs-Lab presents an interactive benchmark for studying AI-driven requirement elicitation. An assistant iteratively refines incomplete requirements by analyzing datasets and asking clarification questions. The benchmark spans 16 public datasets, user profiles, and an automated evaluation pipeline using AI-simulated users to assess elicitation quality.
Academic testbed for studying how AI assistants convert vague goals into actionable requirements. Relevant to researchers building elicitation tools; limited immediate utility for practitioners.
Provenance-aware pipeline merges system telemetry and human-authored reports into a shared typed task-state representation for human-AI task handover. Evaluation on 13 paired task states in a controlled spatial environment shows structured reconciliation preserves more task-state utility than either source alone, and incurs less misinformation than an end-to-end LLM given the same inputs.
Empirical comparison showing raw LLM handover introduces measurable misinformation; a design pattern worth borrowing when delegating state to agents.
Research paper comparing ML and DL models for classifying balanced versus imbalanced postural states in VR using kinematic, EMG, and EDA signals. A Mamba-inspired CNN reached 96.76% accuracy; SHAP analysis showed kinematic features dominated and that a 33% feature reduction preserved performance.
Niche VR balance-detection study with code release; useful as a reference for multimodal posture sensing but only tangentially relevant to conventional desk ergonomics.
An interview study of 12 students examines how they delegate professional emails to generative AI, finding that AI involvement removes the writing skill learning loop, decouples messages from their specific recipient, and redirects confidence from the writer to the system. Two risks are identified: stunted individual capacities and added labor of managing authenticity.
Worth reading for its side-by-side model of AI-mediated versus unaided writing, and the specific framing of who captures the returns from cognitive delegation.
Academic paper proposing a preserved evidence-bundle model and policy-controlled offline verifier for cross-organization agent-to-agent workflows. Each bundle carries authentication, log commitment, witness-backed checkpoints, continuity proofs, delegation authorization, and signed receipts. Prototype tested on 300 workflows; corrupted bundles correctly rejected.
Addresses a concrete gap in multi-agent accountability: offline, policy-controlled verification without trusting live systems. Useful reference for anyone building auditable cross-organization agent pipelines.
FedEHR-Agents proposes a federated framework where autonomous clinical LLM agents at different hospitals collaborate by sharing modeling experience and meta-prompts rather than patient data, using TextGrad-based prompt refinement and evidence-guided aggregation to improve automated EHR modeling across multi-hospital benchmarks.
Frames experience and prompt sharing as a federation object for multi-agent systems under privacy constraints, a transferable pattern beyond the clinical setting.
A position paper arguing that human-centered explainable AI should incorporate psychology of information-seeking. It proposes users evaluate explanations via instrumental, hedonic, and cognitive expected utilities, shaped by cognitive biases, leading to either excessive or insufficient information-seeking with agentic AI systems.
Provides a framework for designers of agentic tools to move from making explanations available to making them sought, grounded in documented cognitive biases rather than assumptions about user behavior.
An arXiv paper examines AI-assisted sensemaking of qualitative UX data through a case study of 20 user responses on video conferencing platforms. It proposes a framework positioning AI as an amplifier of human interpretive judgment rather than a replacement, identifying when algorithmic efficiency helps versus hinders designer agency.
Provides a concrete framework and case study for integrating AI into qualitative research workflows without losing interpretive depth. Useful for UX practitioners and researchers designing AI-assisted analysis pipelines.
Proposes the SCAN framework (Substitute, Complement, Aid, Non-Negotiable) for allocating clinical reasoning tasks between medical learners and generative AI. Reframes learner difficulty as misclassification of interaction mode rather than misuse, and identifies passive engagement as a detection-resistant path to mis-skilling.
Names a concrete failure pattern (passive engagement in correctly scaffolded tasks) and supplies a testable decision vocabulary. Useful for educators and curriculum designers, though the framework is conceptual and unvalidated.
A mixed-methods user study examining how people choose between world-anchored and body-anchored mixed reality interface elements across stationary and mobile contexts, finding anchoring preferences shift with mobility and depend on personal factors like accessibility, stability, and visual clutter.
Empirical data on MR anchoring trade-offs that can guide adaptive interface design for spatial computing workflows.
The paper shows that Learning to Defer (LtD) strategies create class-dependent sampling bias, disproportionately routing minority-class items to human experts. A user study (N=226) finds that imbalanced deferred item sets trigger human cognitive bias, reducing accuracy in majority-class classification.
Concrete evidence that how an AI chooses to defer decisions can distort human judgment downstream, relevant for designing human-AI collaboration workflows.
A user study (N=204) introduces Graphionale, a system that converts linear LLM rationales into interactive multi-level argument graphs. Graphical rationales improved trust calibration for verbal reasoning but felt more demanding; for visual reasoning, they impaired calibration yet felt more engaging and helpful.
Empirical rule for matching rationale format to task modality, with a working testbed and a 204-participant study behind the claim.
Research paper drawing on two studies of experienced Wikipedia editors that proposes an AI-Assisted Sensemaking Model for collaborative discussions, capturing arguments, group norms, participants, and shared history, with a spectrum of interpretive support from low to high.
Identifies a real gap: most AI discussion tools ignore the social context that gives group conversations meaning. The model and its tradeoff analysis are directly useful for anyone designing AI-mediated collaboration tools.
MemToC is a controlled benchmark for evaluating how LLM agents arbitrate between parametric memory and tool returns when the two conflict. Across five 7-9B open-weight models, instruction-tuned variants retain a verified-correct answer against an incorrect tool in only 6.5-17.1% of cases. SFT and DPO improve arbitration on some backbones but reduce abstention after tool errors.
Quantifies a specific failure mode in tool-using agents and offers a benchmark plus fine-tuning baseline for anyone designing agent arbitration policies.
