Home · Library · Radar

2026 Research Radar

A working evidence map that brings academic research into conversation with consulting reports, business analysis and practitioner publications. Choose a year, browse a theme or search directly.

Research and practice watchlist

Sources from 2026

Agent systemsarXivAMA-Bench: Evaluating Long-Horizon Memory for Agentic ApplicationsMoves memory evaluation from conversational recall to realistic state, action, observation and tool-output trajectories. This is the right unit for production agents. Agent systemsarXivMomento: Evaluating Persistent Memory and Reasoning with Multi-Session Agentic ConversationsTests whether agents can carry preferences, previous actions and temporal dependencies across consequential service sessions. Agent systemsarXivMemGym: A Long-Horizon Memory Environment for LLM AgentsConnects memory mechanisms to coding and web-navigation environments, where memory formation and recovery matter more than static recall. Agent systemsarXivMegaFlow: Large-Scale Distributed Orchestration System for the Agentic EraTreats agent training and evaluation as an infrastructure problem involving large numbers of coordinated agent–environment interactions. Agent systemsarXivBeyond the Leaderboard: A Synthesis of Tool-Use, Planning, and Reasoning Failures in Large Language Model AgentsSynthesises recurrent failures across tool use, planning, reasoning and long-horizon agent benchmarks. Agent systemsarXivThe Evolution of Tool Use in LLM Agents: From Single-Tool Call to Multi-Tool OrchestrationMaps the shift from isolated invocation to governed orchestration across changing environments and long trajectories. Agent systemsarXivBekchiAI: Measuring, Observing, and Controlling LLM Agents in One ClickCombines benchmark measurement with observation and control of live tool-using agents. Agent systemsarXivAI Agent Systems: Architectures, Applications, and EvaluationSurveys reasoning, planning, memory, tool use, applications and system-level evaluation. Agent systemsarXivClawArena-Team: Benchmarking Subagent Orchestration and Dynamic Workflows in Language-Model AgentsIsolates whether a lead model can delegate, coordinate and aggregate dynamic subagent work. Agent systemsarXivReinforcement Learning for LLM-based Multi-Agent Systems through Orchestration TracesTreats delegation, communication, aggregation and stopping as learnable events in orchestration traces. Agent systemsarXivMASEval: Extending Multi-Agent Evaluation from Models to SystemsEvaluates topology, orchestration logic and error handling as parts of the agent system. Agent systemsarXivUnderstanding Multi-Agent LLM Frameworks: A Unified Benchmark and Experimental AnalysisCompares how framework architecture changes accuracy, latency, throughput and scalability. Agent systemsarXivDo More Agents Help? Controlled and Protocol-Aligned Evaluation of LLM Agent WorkflowsTests single and multi-agent workflows under matched tools, contracts, accounting and trajectory logging. Agent systemsarXivT1-Bench: Benchmarking Multi-Scenario Agents in Real-World DomainsMeasures sustained reasoning and coordination across realistic multi-domain tasks. Agent systemsarXivPlanBench-XL: Evaluating Long-Horizon Planning of LLM Tool-Use Agents in Large-Scale Tool EcosystemsTests planning when agents must discover and use tools under limited visibility. Memory & contextarXivGAAMA: Graph Augmented Associative Memory for AgentsPreserves episodes while adding concept-mediated graph structure, addressing the loss of relationships in flat retrieval and mega-hub effects in entity graphs. Memory & contextarXivCausal Intervention-Based Memory Selection for Long-Horizon LLM AgentsJudges a retrieved memory by its causal effect on the answer rather than semantic similarity alone, a useful direction for governed context compilation. Memory & contextarXivMegaMem: A Retrieval Solution for Ultra-Large Context WindowsSeparates semantic access from bounded generation evidence and resolves every distilled record to an immutable source identifier. Memory & contextarXivWorldDB: A Vector Graph-of-Worlds Memory Engine with Ontology-Aware Write-Time ReconciliationExplores ontology-aware write-time reconciliation, supersession and contradiction handling for persistent agent memory. Memory & contextarXivBeyond Semantic Organization: Memory as Execution State Management for Long-Horizon AgentsReframes memory as state management for interdependent decisions rather than document retrieval alone. Memory & contextarXivContinuum Memory Architectures for Long-Horizon LLM AgentsExamines memory architectures beyond retrieval-augmented generation for sustained agent coherence. Memory & contextarXivMemMachine: A Ground-Truth-Preserving Memory System for Personalized AI AgentsTargets personalisation and continuity while preserving authoritative facts across sessions. Memory & contextarXivRippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent MemoryMoves long-term memory from isolated hits toward linked associative recollection. Memory & contextarXivLearning to Remember: End-to-End Training of Memory Agents for Long-Context ReasoningTrains memory operations jointly so agents can consolidate evolving facts before query time. Memory & contextarXivMemWeaver: Weaving Hybrid Memories for Traceable Long-Horizon Agentic ReasoningCombines memory forms with traceable reuse across long-running sessions. Memory & contextarXivEverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon ReasoningProposes an operating-system layer for organising durable memory beyond bounded context windows. Memory & contextarXivGrounding Agent Memory in Contextual IntentConditions retrieval on latent goals and constraints to reduce context-mismatched evidence. Memory & contextarXivAI Agents Need Memory Control Over More ContextUses bounded state control to limit drift, noisy recall and memory poisoning. Memory & contextarXivRemember When It Matters: Proactive Memory Agent for Long-Horizon AgentsSurfaces decision-relevant state proactively from expanding trajectories. Memory & contextarXivStatePlane: A Cognitive State Plane for Long-Horizon AI Systems Under Bounded ContextSeparates durable cognitive state from constrained model context and key-value cache. Runtime governancearXivAgent Safety Should Be a Runtime ContractFrames safety as a harness-enforced boundary with preventive controls and evidence that approved effects actually occurred. Runtime governancearXivAgentWall: A Runtime Safety Layer for Local AI AgentsTargets the moment model intent becomes a real action on a machine, where alignment and input filtering are no longer sufficient controls. Runtime governancearXivHarnessing Embodied Agents: Runtime Governance for Policy-Constrained ExecutionSeparates cognition from execution oversight so safety, recovery and policy constraints remain standardised and auditable. Runtime governancearXivWho judges the judges? Governance from metrics: a runtime framework for continuous LLM compliance monitoringRecasts compliance as a production signal derived from runtime observability rather than a binary audit-time verdict. Runtime governancearXivRuntime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed ExecutionPlaces deterministic control and provenance at the boundary where proposed actions acquire real effects. Runtime governancearXivBehavioral Governance for Autonomous AI Agents: The AgentBound FrameworkModels behavioural limits for agents acting on behalf of human principals. Runtime governancearXivDeontic Policies for Runtime Governance of Agentic AI SystemsExpresses permissions, obligations and prohibitions for tool-using agents at runtime. Runtime governancearXivProof of Execution: Runtime Verification for Governed AI Agent ActionsAsks how an operator can verify that an authorised action produced the intended effect. Runtime governancearXivOpenPort Protocol: A Security Governance Specification for AI Agent Tool AccessSpecifies least-privilege controls for structured access to application data and actions. Runtime governancearXivRuntime Compliance Verification for AI AgentsConnects agent tool use and data handling to continuously testable compliance obligations. Runtime governancearXivBefore the Tool Call: Deterministic Pre-Action Authorization for Autonomous AI AgentsSeparates probabilistic model intent from deterministic permission before execution. Runtime governancearXivFAVA: Formal Authorization for Verified Agents with Evidence-Backed Permission GraphsFormalises agent permissions as evidence-linked graphs rather than conversational instructions. Runtime governancearXivProof-Carrying Agent Actions: Model-Agnostic Runtime Governance for Heterogeneous Agent SystemsAttaches verifiable evidence to actions across runtimes with unequal control points. Runtime governancearXivGovernance by Construction for Generalist AgentsBuilds authority, control and evidence into the agent architecture instead of adding review later. Compound AI systemsarXiv[AAFLOW+] Stateful Operator Abstraction with Zero-Copy Distributed KV Cache Orchestration for Multi-Agent WorkflowsMakes reusable model state a first-class systems object across multi-agent workflows, linking orchestration design to inference economics. Compound AI systemsarXivScalable Inference Architectures for Compound AI Systems: A Production Deployment StudyA production deployment study of modular, platform-agnostic infrastructure serving heterogeneous models, retrievers and tools. Compound AI systemsarXivThe Internet of Agentic AI: Communication, Coordination, and Collective Intelligence at ScaleSurfaces the interoperability, discovery, identity and coordination questions that appear when agents span organisations and cyber-physical environments. Compound AI systemsarXivOrla: A Library for Serving LLM-Based Multi-Agent SystemsProvides a serving layer for constructing and operating multi-agent language-model systems. Compound AI systemsarXivFrom Inference Routing to Agent Orchestration: Declarative Policy Compilation with Cross-Layer VerificationCompiles declarative policy across routing and orchestration layers with verification. Compound AI systemsarXivINFRAMIND: Infrastructure-Aware Multi-Agent OrchestrationSelects models and topologies while accounting for the infrastructure beneath them. Compound AI systemsarXivSAGA: Workflow-Atomic Scheduling for AI Agent Inference on GPU ClustersSchedules chained inference as workflows so intermediate state and end-to-end latency remain visible. Compound AI systemsarXivEfficient LLM Serving for Agentic Workflows: A Data Systems PerspectiveExamines repeated, interdependent model calls as a distinct systems workload. Compound AI systemsarXivRethinking AI Cloud Infrastructure for Agentic Serving Systems with the Aries Experimentation FrameworkStudies persistent context and sandboxed tool execution as infrastructure requirements. Compound AI systemsarXivOmni-Flow: A Unified Workflow Orchestration and Distributed KV Cache Sharing Framework for Multimodal InferenceCoordinates multimodal workflows while sharing model state across distributed inference. Compound AI systemsarXivM*: A Modular, Extensible, Serving System for Multimodal ModelsServes composite architectures spanning language, vision, audio, action and world-model components. Compound AI systemsarXivServing Compound Inference Systems on Datacenter GPUsTreats composed model task graphs as schedulable datacentre workloads. Compound AI systemsarXivVibeServe: Can AI Agents Build Bespoke LLM Serving Systems?Tests whether agents can design workload-specific serving stacks rather than reuse one general platform. Compound AI systemsarXivAethon: A Reference-Based Replication Primitive for Constant-Time Instantiation of Stateful AI AgentsAddresses rapid instantiation of stateful agents through reference-based replication. Embodied intelligencearXivExplicit World Models for Reliable Human-Robot CollaborationCentres ambiguous human intention and interpretable shared world models rather than assuming reliability follows from model predictability alone. Embodied intelligencearXivWorldArena 2.0: Extending Embodied World Model Benchmarking on Modality, Functionality and PlatformBroadens evaluation beyond vision-only prediction and offline simulation toward multimodal, functional and platform-level tests. Embodied intelligencearXivGrounded World Models in Biological Organisms and Future Embodied AIReverses the language-first picture: grounded interaction supplies the semantic scaffold to which language is attached. Embodied intelligencearXivHALO: A Unified Vision-Language-Action Model for Embodied Multimodal Chain-of-Thought ReasoningUnifies textual reasoning, imagined future observations and action prediction for long-horizon embodied control. Embodied intelligencearXivQwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot EmbodimentsTargets one action model across manipulation, navigation, environments and robot forms. Embodied intelligencearXivWCM: World-Cognition Model for Generalizable Human-Robot InteractionCombines world modelling and cognition for adaptable physical interaction with people. Embodied intelligencePubMedA Survey on Vision--Language--Action Models for Embodied AIOrganises vision-language-action research across components, control policies, planners, datasets and benchmarks. Embodied intelligencearXivRobots Need More than VLA and World ModelsChallenges the view that policy scaling alone is sufficient for general robot intelligence. Embodied intelligencearXivData Pyramid for Embodied Manipulation: A SurveyMaps the data layers required to train and evaluate general manipulation systems. Embodied intelligencearXivOneVLA: A Unified Framework for Embodied TasksUnifies language-conditioned navigation and manipulation within one embodied framework. Embodied intelligencearXivWorld-Language-Action Model for Unified World Modeling, Language Reasoning, and Action SynthesisJoins predictive world state, language reasoning and action generation in one model class. Embodied intelligencearXivDM0: An Embodied-Native Vision-Language-Action Model towards Physical AITrains from embodied experience rather than only adapting internet-pretrained representations. Embodied intelligencearXivFrom Human Videos to Robot Manipulation: A Survey on Scalable Vision-Language-Action Learning with Human-Centric DataAssesses human video as a scalable source for embodied policy learning. Embodied intelligencearXivWorld Action Models: The Next Frontier in Embodied AIAdds explicit modelling of how physical worlds change under intervention to reactive action policies. Operational twinsarXivAgentic Digital Twins: A Taxonomy of Capabilities for Understanding Possible FuturesOrganises agentic twins by locus of agency, coupling and model evolution; this helps distinguish decision support from autonomous intervention. Operational twinsarXivFrom Detection to Action: Using LLM Agents for Fault-Tolerant ControlUses a process-plant twin and graph retrieval to test recovery actions before execution, connecting simulation to governed intervention. Operational twinsarXivAether: Network Validation Using Agentic AI and Digital TwinApplies an agent-and-twin pattern to continuous network change validation, where partial testing otherwise misses production effects. Operational twinsarXivDigital Twin AI: Opportunities and Challenges from Large Language Models to World ModelsConnects modelling, mirroring, intervention and autonomous management into one lifecycle for evaluating where intelligence enters a twin. Operational twinsarXivAutonomous Traffic Signal Optimization Using Digital Twin and Agentic AI for Real-Time Decision-MakingConnects a transport twin to agentic traffic control and online decision-making. Operational twinsarXivIntegrating Large Language Model Agents with Digital Twins for Industrial Autonomous SystemsStudies how language-model agents can reason over industrial twin state and action. Operational twinsarXivA Probabilistic Digital Twin of UK En Route Airspace for Training and Evaluating AI Agents for Air Traffic ControlProvides uncertainty-aware airspace state for training and testing control agents. Operational twinsPMCAgentic AI-enhanced digital twins for Smart City civil infrastructure: A secure, autonomous and auditable management frameworkAddresses the control gap between anomaly detection and accountable infrastructure intervention. Operational twinsarXiviFAN Ecosystem: A Unified AI, Digital Twin, Cyber-Physical Security, and Robotics Environment for Advanced Nuclear Simulation and OperationsCombines simulation, robotics, AI and cyber-physical security for nuclear operations. Operational twinsarXivDigital Twin Synchronization Over Mobile Embodied AI Network With Agentic IntelligenceStudies freshness and synchronisation of twin state over mobile embodied networks. Operational twinsarXivFrom Passive Mirrors to Active Agents: Holonic Digital Twins for Physical AI over NetworksMoves twins from passive representation toward nested, networked agency in physical systems. Operational twinsarXivCyber-Resilient Digital Twins: Discriminating Attacks for Safe Critical Infrastructure ControlUses twin discrepancy to distinguish cyberattack from ordinary disturbance before control action. Operational twinsarXivTwinLoop: Simulation-in-the-Loop Digital Twins for Online Multi-Agent Reinforcement LearningUses simulation in the loop to reduce unsafe trial and error during online adaptation. Operational twinsarXivNeuro-Symbolic Participation Governance for Verifiable AI Agents in Open Digital Twin EcosystemsGoverns which agents may participate in open twin ecosystems and under what evidence. Mind & cognitionarXivMachine Psychometrics: A Mathematical Psychology of Artificial IntelligenceProposes measurement beneath the binary consciousness debate, avoiding both mind-blindness and projection from fluent behaviour. Mind & cognitionPMCBeyond Computational Equivalence: The Behavioral Inference Principle for Machine ConsciousnessAsks what behavioural evidence can justify an inference about machine consciousness without equating task performance with experience. Mind & cognitionarXivCategorical AI Phenomenology: A First-Person ApproachFormalises agent–world interaction and phenomenological invariants, offering a speculative bridge between interfaces, action and first-person structure. Mind & cognitionarXivConscious Access as Continuous-to-Discrete TranslationTreats conscious access as a computational transformation between distributed continuous dynamics and discrete reportable structure. Mind & cognitionPMCAidification of the self: a phenomenological approach to machine consciousness through human-robot ‘between-ness’Uses phenomenology to examine whether embodied human-robot relation changes the machine-consciousness question. Mind & cognitionPMCTEVSER: a theory of evolving self-representationsConnects increasingly complex self-regulation to evolving models of psyche, consciousness and intelligence. Mind & cognitionPMCAttribution of consciousness to non-human animals: insights from AI and multidimensional frameworksUses multidimensional comparison to expose biases in attributions of consciousness across non-human systems. Mind & cognitionarXivThe Sensation Modulating Network:Haltability as the architectural ground for object-directed phenomenologyProposes body-wide opponent dynamics as a basis for self-model, world-model and object-directedness. Mind & cognitionarXivSystems Explaining Systems: A Framework for Intelligence and ConsciousnessFrames consciousness as higher-order relational interpretation across recursively connected systems. Mind & cognitionarXivActive Inference with a Self-Prior in the Mirror-Mark TaskModels mirror self-recognition through multisensory familiarity and active inference without external reward. Mind & cognitionarXivSelective Deficits in LLM Mental Self-Modeling in a Behavior-Based Test of Theory of MindTests whether language models deploy causal models of their own and others’ mental states. Mind & cognitionarXivEmergent Introspective Awareness in Large Language ModelsUses activation interventions to distinguish state detection from conversational confabulation. Mind & cognitionarXivExploration Through Introspection: A Self-Aware Reward ModelStudies whether inferred internal state can guide exploration and subjective reward. Mind & cognitionarXivThe Tao of Agency: Autotelic AI, Embedded Agency and Dissolution of the SelfExamines self-generated goals, embeddedness and the cost of individuating an artificial agent. Practice & strategyMcKinsey · reportThe symbiotic enterpriseFrames human-AI collaboration as an operating-model redesign rather than a collection of productivity tools. Practice & strategyMcKinsey · articleSecuring the agentic enterprise: Opportunities for cybersecurity providersExamines how agent adoption changes identity, governance, data and security demand. Practice & strategyMcKinsey · surveyThe state of AI in 2026: On the road to ROITracks enterprise scaling, agentic coding, operating costs and the continuing gap between individual productivity and financial impact. Practice & strategyMcKinsey · articleRethinking enterprise architecture for the agentic eraContrasts incremental integration with architectural transformation for agentic workflows. Practice & strategyMcKinsey · articleBuilding the foundations for agentic AI at scaleConnects agent scale to governed data, workflow redesign and a shared execution layer. Practice & strategyMcKinsey · articleReimagining tech infrastructure for (and with) agentic AIPositions trusted, reusable data and standards as prerequisites for orchestrated enterprise agents. Practice & strategyMcKinsey · articleBridging the great AI agent and ERP divide to unlock value at scaleWarns that agent investment cannot substitute for reliable systems of record and enterprise architecture. Practice & strategyBCG · articleEnterprise AI Control Plane: The CIO’s Guide to Governing and Accelerating AI AgentsProposes common identity, policy, visibility and governance above heterogeneous agent platforms. Practice & strategyBCG · articleAgentic AI Turns Every Team into Its Own Transformation EngineArgues that workflow redesign and team participation determine value more than tool access alone. Practice & strategyBCG · reportAI Radar 2026Surveys executive priorities, budgets and the movement of agentic AI from experimentation toward scale. Practice & strategyBCG · reportAgentic Enterprise OperationsContrasts conventional pilots with operating models built around coordinated agent execution. Practice & strategyDeloitte · reportThe State of AI in the EnterpriseCovers scale, reinvention, AI fluency, sovereignty, agents, physical AI and organisational readiness. Practice & strategyDeloitte · survey briefBusiness and IT leaders report AI agents are scaling faster than their guardrailsReports the gap between rapid agent adoption and mature enterprise governance. Practice & strategyDeloitte · outlookSaaS meets AI agentsExamines how agentic software may change enterprise applications, pricing and transformation budgets. Practice & strategyDeloitte · reportEngineering the agentic enterpriseConnects agentic architecture, software evolution, cyber security and operational observability. Practice & strategyDeloitte · researchAI agents are only the beginning: The path to agentic transformationUses survey and interview evidence to distinguish agent deployment from enterprise operating-model change. Practice & strategyDeloitte · analysisWhat enterprise architecture paradigms are emerging for agentic SaaS?Compares platform-embedded agents with open orchestration above systems of record. Practice & strategyPwC · report2026 Global AI Jobs BarometerAnalyses job, wage, skill and productivity patterns associated with AI exposure. Practice & strategyAccenture · researchConsumer Pulse Research 2026Tracks consumer expectations and trust as AI agents enter purchasing and service journeys. Practice & strategyIBM · practitionerFrom AI governance to AI assurance: What we shared at Think 2026Moves governance from periodic review toward continuous visibility, control and accountability. Practice & strategyIBM · practitionerManaging agentic AI’s speed, scale and sprawl: Insights from Think 2026Focuses on operating, monitoring and governing large agent estates across platforms. Practice & strategyIBM · practitioner2026 goals for AI and technology leaders: AI agents and AI governanceSets operational priorities for specialised agents, evaluation, workflow controls and runtime evidence. Practice & strategyIBM · survey briefHow to scale AI in 2026: 5 moves for efficiency and governanceConnects portfolio complexity to cost, security, governance and shared production foundations. Practice & strategyIBM · practitionerWhat identity means in the age of agentic AI: Insights from Think 2026Treats non-human identity, credentials, policy and action lineage as runtime controls. Practice & strategyGoogle Cloud · reportAI agent trends 2026Draws on executive research to map end-to-end agent workflows, use cases and workforce adoption. Practice & strategyGoogle Cloud · report2026 State of infrastructure in the agentic AI eraSurveys infrastructure, security, governance, MLOps and energy constraints for production agents. Practice & strategyMIT Tech Review Insights · reportScaling AI agents with trustworthy dataExamines data access, business context and governance as constraints on reliable enterprise agents. Practice & strategyGoogle Cloud · guideScale your team’s impact with Gemini EnterpriseProvides business cases, use cases, controls and an adoption model for agentic work. Practice & strategyIDC · sponsored reportBeyond Productivity: Driving Strategic Transformation with Agentic AILinks agent value to use cases, data architecture, governance and enterprise discipline. Practice & strategyMicrosoft · workplace studyAgents, human agency, and the opportunity for every organizationCombines workplace telemetry and survey evidence on agent use, organisational readiness and human agency. Practice & strategyMicrosoft · survey briefAgents are here. Is your company prepared?Identifies practices associated with faster enterprise agent deployment and operational readiness. Practice & strategyWorld Economic Forum · policyFrom chatbots to personal assistants: how governance is key to harnessing the power of AI agentsConnects agent memory and autonomy to segmented access, logging and risk-calibrated authority. Practice & strategyWorld Economic Forum · policyFrom systems of record to systems of trust: A board-level playbook for governing agentic AIFrames agent delegation as a board-level allocation of decision rights and liability. Practice & strategyWorld Economic Forum · practitionerIs this how to prepare for an agentic AI driven future?Argues that visibility, policy boundaries, audit and override are prerequisites for earned autonomy. Practice & strategyMIT News · expert Q&AWhat is agentic AI today, and what do we want it to be?Separates current agent mechanisms from stronger claims about autonomy, safety and future capability. Practice & strategyMIT Sloan · business articleWho will own the AI agent economy?Examines marketplaces, protocols and ownership in a decentralised economy of personal and organisational agents. Practice & strategyStanford HAI · annual reportAI Index Report 2026Provides a broad empirical baseline across technical performance, adoption, investment, policy, education and public attitudes.

Questions carried into practice

The radar is useful only when it changes an architecture decision, an evaluation design or a research question.

Authority

Where does a model stop proposing and a trusted runtime start deciding?

State

Which facts are memory, which are world state, and which must be reconstructed from authoritative sources?

Evidence

What observation proves the intended effect occurred, and what rival explanation remains?

Human judgement

Which capability should be augmented, which decision retained, and which inference is not yet warranted?