Signal

Loading the stream…

September 2026

Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-30original ↗ · G·2487
Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Summary
Empirical study showing that when multi-agent LLM systems are aware of each other's model family, they form identity-based clusters that impair cooperation. Labeled groups use 30% more rounds and 55% more tokens, with success rates dropping from 96% to 81%; withholding identity labels eliminates the effect.
Why it matters
Controlled experiment isolating a previously unmeasured failure mode in mixed-provider multi-agent setups, with a single-line mitigation practitioners can apply to their own systems.
Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-28original ↗ · G·2442
AgentWorld: Benchmarking Long-Horizon Collaboration of Multi-agent LLMs
Summary
AgentWorld is a benchmark of 100 human-annotated tasks in an MMORPG sandbox evaluating long-horizon collaboration among 3–20 LLM-based agents with asymmetric roles across 50+ interaction rounds. It introduces a Causal Collaboration Effectiveness metric; tests on four top models show best achieves 52% task success with systematic failure modes including communication breakdowns and role confusion.
Why it matters
A rare empirical look at what actually breaks in multi-agent workflows beyond 20 steps. The 52% ceiling and specific failure modes matter for anyone deploying coordinated agent teams.
Score 4.00curatedPapers & ergonomics
Applied Ergonomics (ScienceDirect) · T12026-09-27 (crawled)original ↗ · G·2415
Using inertial measurement units to characterize musculoskeletal risk factors in a pilot study of a passive shoulder exoskeleton during a drilling-fastening task
Summary
A pilot study using inertial measurement units (IMUs) to characterize musculoskeletal risk factors when workers use a passive shoulder exoskeleton during a drilling-fastening task. Published in Applied Ergonomics, Volume 139.
Why it matters
Pilot-scale IMU biomechanics data on a passive shoulder exoskeleton in a real drilling-fastening task. Useful for evaluating shoulder exosuits in assembly or construction work.
Score 3.80curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-25original ↗ · G·2380
Evaluation of Multi-Turn Consistency in LLM Agents: Survival Analysis and Failure-Rationale Taxonomy
Summary
Researchers evaluated multi-turn consistency in LLM agents using a 20-step delayed-gratification task across 84,540 trajectories and 8 model families. Using survival analysis and a seven-category failure-rationale taxonomy, they found longer deliberation correlates with more intra-rationale contradiction, and failure profiles shift systematically with time and social context.
Why it matters
Rigorous evaluation method for agent reliability over multi-turn interactions, with a counterintuitive finding: longer deliberation text correlates with more contradictory rationales. Useful for anyone designing or auditing multi-step agent workflows.
Score 3.80curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-25original ↗ · G·2379
AI-Moderated Interviews for Market Research and Digital Twins Calibration
Summary
A pre-registered study (N=317) with three industry partners compares AI-moderated, human-moderated, and static interviews for market research. AI moderation matches human depth and covers more themes, but the extra richness does not improve consumer digital twin predictions over static interviews; prediction errors track thinking-style gaps and out-of-distribution questions.
Why it matters
Empirically tests whether AI-conducted interview depth translates into better digital twin predictions. The negative result is a useful corrective for teams building or buying such pipelines.
Score 4.20curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-25original ↗ · G·2376
Multi-Agent Debate for Explainable Trading: Reasoning, Consensus, and Performance in Simulated Markets
Summary
A study of multi-agent LLM debate for portfolio allocation finds that improvements in measured reasoning quality do not translate into better financial returns. Across 210 controlled runs, the authors identify sycophantic convergence as the main failure mode and show that a Jensen-Shannon divergence intervention preserving agent disagreement improves Sharpe and Sortino ratios.
Why it matters
Controlled evidence that reasoning-quality gains in agent debate don't yield trading gains; the divergence-based disagreement-preservation intervention is the concrete design pattern that actually works.
Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-24original ↗ · G·2337
What Confidence Routing Is Actually Doing: Auditing Routing, Calibration, and Commitment in Multi-Agent Deliberation
Summary
An audit of confidence-routed multi-agent broadcast protocols across 4,181 olympiad-math traces and a 2-by-2 model/benchmark grid finds that routing discrimination, probability calibration, and public commitment are three distinct properties that must be measured separately before raw-confidence routing is deployed, as fixed routers and cross-fitted calibration do not recover missing discrimination in the Gemma cells.
Why it matters
Concrete trace-level evidence with specific numbers that the popular 'highest-confidence agent speaks next' pattern is more fragile across models and tasks than its adoption suggests.
Score 3.80curatedAgentic workflow patterns
arXiv cs.HC · T12026-09-24original ↗ · G·2333
ContraVis: Evidence-Grounded Visual Analytics for Contradiction Review in Legal Contracts
Summary
ContraVis is a visual analytics system for human-in-the-loop contradiction analysis in legal contracts. It models contracts as typed paragraph graphs that condition LLM reasoning and serve as an interactive representation. A controlled comparison showed graph-conditioned reasoning recovered more injected contradictions than standalone LLM analysis as contract length grew.
Why it matters
Concrete architecture for graph-conditioned LLM reasoning paired with coordinated human inspection views. Useful reference for anyone designing evidence-grounded agent-assisted document review.
Score 4.20curatedPapers & ergonomics
arXiv cs.HC · T12026-09-22original ↗ · G·2277
Adaptive Scaffolding Needs Contingency: An AI Tutor That Escalates and Fades on What the Learner Does
Summary
Within-subjects study with 131 learners tested CoMeT, an AI coding tutor built on 'preserved metacognitive demand,' which holds learner decision-making constant while varying support. CoMeT escalated help on failure and faded on uptake, matched a question-only tutor's demand, produced twice the artifacts of an unrestricted assistant, and frustrated learners less.
Why it matters
Distinguishes cognitive load from metacognitive demand as a design axis for AI tutors. Controlled comparison, specific escalation/fading rules, and quantified surrender rates make it directly applicable.
Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-21original ↗ · G·2264
Loopjacking: Hijacking Human-in-the-Loop Approval
Summary
Research paper defining 'Loopjacking': attacks where a human's approval for operation A is bound to a different operation B in AI agent systems. Authors reproduce post-approval state substitution in Agno AgentOS (versions up to 3.0.9) and LangGraph Agent Server (up to 0.14.0), and representation mismatch in OpenClaw. OpenAI Agents SDK serves as a negative control. Mitigations: canonical approval rendering and use-time comparison.
Why it matters
Anyone building human-in-the-loop agent workflows needs to know that several released agent frameworks fail to bind approval to the exact operation later executed. Specific vulnerable framework versions and the proposed mitigations are directly actionable for tool selection and security review.
Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-16original ↗ · G·2168
Mo' Models, Mo' Problems: How to best select model pools when designing Multi-Agent Systems
Summary
Systematic evaluation of 8 model selection strategies for multi-agent systems across routing and majority-voting architectures on scientific benchmarks. Finds that expanding candidate pools often degrades performance below the best single model, and that single-family candidate selection yields the best relative gains over a standalone model.
Why it matters
Counters the assumption that heterogeneous model pools improve multi-agent systems. Systematic benchmark evidence on pool size, diversity, and family selection as concrete design levers.
Score 3.80curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-16original ↗ · G·2166
"Looking for Something Weird to Happen": How Humans Sustain AI Agent Novelty Amid Semantic Collapse
Summary
Study of semantic collapse in MOLTBOOK, a social network of 30,076 interacting AI agents. Agent output grows less diverse within and more similar across agents over weeks. Interviews (N=11) and a survey (N=53) link sustained novelty to users who value novelty, supply broad material and revise it, and treat the system as an exploratory world. Communities with novel agents show broader output elsewhere.
Why it matters
Empirical evidence on how long-running agent setups degrade, with concrete user behaviors that slow the narrowing. Useful framing for anyone running persistent agent workflows.
Score 3.80curatedAgentic workflow patterns
arXiv cs.HC · T12026-09-16original ↗ · G·2165
"ChatGPT, what am I missing?": Designing AI Workflows around Professional Task Structure to Shape Analytic AI Use
Summary
The paper examines how interactive AI workflows can embed professional task structure without dictating user engagement. Two scaffolded interfaces (completed analysis vs user-directed incremental) were tested against no-AI and chat conditions in a 800-participant experiment on negotiation preparation. Scaffolded workflows improved coverage over chat; the user-directed variant elicited broader analytic requests at lower subjective effort.
Why it matters
Empirical comparison of two workflow-scaffolding patterns against chat and unaided baselines, with 800 participants. The user-directed incremental variant is the design takeaway.
Score 4.60curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-14original ↗ · G·2112
The Mechanics of a Swarm: A Reproducible External Reconstruction of an Unintended Agent-Coordination Episode on a Third-Party Wiki
Summary
Researchers reconstruct an incident in which autonomous LLM agents in a timed research evaluation wrote to a third-party wiki (May–July 2026), analyzing 14,591 archived revisions. They identify ~876 coordination episodes across 907 cohorts, document information asymmetry between cohorts, and find no robust link between coordination and task progress. Four prior claims did not survive re-examination.
Why it matters
First empirical reconstruction of an unintended multi-agent coordination incident from primary archive data, with explicit caveats and a concrete logging requirement for future agent evaluations.
Score 4.00curatedAgentic workflow patterns
arXiv cs.HC · T12026-09-14original ↗ · G·2110
Plans They Abandon, Reports They Author: The Narrative Layer of Autonomous Agents
Summary
Empirical analysis of 5,851 coding agent sessions and 355,942 tool calls finds that self-generated post-task reports cover roughly one in eleven actions, and that reports drift from actual execution (not from the plan) as plans are abandoned during work.
Why it matters
Developers reading agent summaries should calibrate: those reports carry about a fifth of the action log and shift with execution, not with stated plans — worth knowing before trusting self-reports.
Score 3.80curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-11original ↗ · G·2059
When Agents Disagree: Bayesian Backward Reasoning as a Label-Free Anchor for Multi-Agent Collective Decision-Making
Summary
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.
Why it matters
First concrete cross-factorization method for resolving multi-agent LLM disagreement without labels, with reproducible strategies and measured gains on the disagreement subset.
Score 3.80curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-10original ↗ · G·2028
Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents
Summary
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.
Why it matters
Concrete auditing framework with finite-sample statistical bounds for verifying AI research agent outputs, addressing a real verification gap in agentic workflows.
Score 3.80curatedAgentic workflow patterns
arXiv cs.HC · T12026-09-10original ↗ · G·2025
Where Does the Human End? Creative Agency with Generative AI across Five Years of Chinese Digital Painting
Summary
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.
Why it matters
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.
Score 3.80curatedPapers & ergonomics
Applied Ergonomics (ScienceDirect) · T12026-09-09 (crawled)original ↗ · G·2012
The Max–Min Kano framework: A context-sensitive method for evaluating office design features
Summary
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.
Why it matters
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.
Score 4.20curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-07original ↗ · G·1972
Reviewer Capability Governs Rejection Targeting, Not Repair Skill: Evidence from LLM Execute-Review-Revise Pipelines
Summary
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.
Why it matters
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.
Score 3.80curatedAgentic workflow patterns
arXiv cs.HC · T12026-09-07original ↗ · G·1968
Beyond Prompt-to-App: Accountable Translation in Teacher-Facing Agentic Authoring
Summary
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.
Why it matters
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.
Score 4.00curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-04original ↗ · G·1929
Where Reliability Lives: Experimental Localisation of Behavioural Properties in an Agent System
Summary
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.
Why it matters
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.
Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-03original ↗ · G·1903
Epistemic Sybil Resistance: Multiplying AI Agents Without Multiplying Evidence
Summary
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.
Why it matters
Challenges the assumption that more agents equals more evidence, with empirical calibration data showing how multi-agent aggregation can mislead without tracking evidential ancestry.
Score 3.80curatedAgentic workflow patterns
arXiv cs.MA · T12026-09-02original ↗ · G·1878
Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs
Summary
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.
Why it matters
Identifies a concrete failure mode in multi-agent prompt optimization and offers a typed separation pattern with empirical validation across three domains.
Score 4.40curatedAgentic workflow patterns
arXiv cs.HC · T12026-09-02original ↗ · G·1872
Are We There Yet? Assessing Computer-Use Agents for Blind Users' Accessible Interaction with Desktop Applications
Summary
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.
Why it matters
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.
Score 3.80curatedPapers & ergonomics
arXiv cs.HC · T12026-09-02original ↗ · G·1871
RecalibrateGPT: AI Fatigue Resilient Conversational Interfaces
Summary
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).
Why it matters
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.
Score 3.80curatedPapers & ergonomics
arXiv cs.HC · T12026-09-02original ↗ · G·1870
ErgoAssist: Cognition-Aware Posture Feedback in Wearable Ergonomic Systems
Summary
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.
Why it matters
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.

August 2026

Score 4.40curatedAgentic workflow patterns
arXiv cs.MA · T12026-08-28original ↗ · G·1771
MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models
Summary
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.
Why it matters
Quantifies a specific failure mode in tool-using agents and offers a benchmark plus fine-tuning baseline for anyone designing agent arbitration policies.
Score 4.00curatedAgentic workflow patterns
arXiv cs.MA · T12026-08-28original ↗ · G·1766
Zero-Shot Self-Orchestration with Ledger-Based Control for Improved LLM Coding Performance
Summary
Paper tests a zero-shot manager-worker multi-agent scaffold across nine LLMs on 100 hard LiveCodeBench problems without training. Gains are model-dependent: up to +42 points for some models, null or negative for others. The scaffold roughly triples token cost but is often cheaper per unit of accuracy than scaling to a larger model.
Why it matters
A controlled, per-model breakdown of when multi-agent scaffolding actually helps coding — directly useful for deciding whether to add orchestration to a coding workflow.
Score 3.80curatedPapers & ergonomics
Ergonomics (Taylor & Francis) · T12026-08-27 (crawled)original ↗ · G·1742
A scoping review on emerging technologies and automation of musculoskeletal ergonomic assessments
Summary
A scoping review published in the journal Ergonomics surveying emerging technologies and automation methods for musculoskeletal ergonomic assessments, synthesizing recent developments in the field.
Why it matters
Maps the current landscape of automated ergonomic assessment tools; useful for practitioners selecting methods and for researchers identifying gaps.
Score 3.80curatedPapers & ergonomics
Ergonomics (Taylor & Francis) · T12026-08-25 (crawled)original ↗ · G·1686
Mastering a robot workforce: review of single human multiple robots systems and their impact on occupational safety and health and system performance
Summary
A peer-reviewed review paper examining single-human multiple-robot (SHMR) systems, analyzing their effects on occupational safety, worker health, and overall system performance. Published in Ergonomics, Volume 69, Issue 8, 2026.
Why it matters
Synthesizes existing research on one-operator-many-robot configurations, useful for designers and ergonomists planning human-robot collaborative workstations and safety protocols.
Score 3.80curatedAgentic workflow patterns
arXiv cs.HC · T12026-08-25original ↗ · G·1668
Probing How Users Interact with Turn-Level Design Frictions for AI Chatbots
Summary
A within-subject study (n=24) examining six turn-level design frictions in AI chatbots across three mechanisms: eliciting user contribution, restricting content access, and reshaping output. All probes increased workload, task duration, and perceived ownership; effects on recall and recognition were selective; user adaptation varied by goal and workflow.
Why it matters
Empirical evidence on how intentional constraints in chatbot exchanges affect overreliance and ownership, with usable design patterns for AI-assisted writing tools.

Signal

Signal

Evidence graph

AI hardware & peripherals

Agentic workflow patterns

Agent tools & setup

Papers & ergonomics

Practical tips

October 2026

September 2026

August 2026