
A Inteligência Artificial Agêntica designa sistemas de IA que buscam objetivos de múltiplas etapas de forma autônoma, iterando por meio de ciclos de percepção, planejamento, uso de ferramentas e autocorreção, sem exigir intervenção humana em cada ponto de decisão — uma mudança estrutural em relação aos modelos de resposta rápida em direção a arquiteturas que decompõem objetivos complexos em subtarefas, executam ações em ambientes externos, avaliam os resultados e revisam os planos de acordo.
A arquitetura agentiva canônica acopla um núcleo de raciocínio baseado em um modelo de linguagem robusto com um registro de ferramentas — navegadores web, interpretadores de código, clientes de API, sistemas de arquivos, interfaces de banco de dados — e uma arquitetura de memória que abrange o contexto de trabalho de curto prazo, registros episódicos de ações passadas e conhecimento de longo prazo recuperado, permitindo que o agente mantenha a busca coerente por seus objetivos em horizontes de interação estendidos que ultrapassam qualquer janela de contexto individual. Configurações multiagentes ampliam ainda mais esse conceito, distribuindo subtarefas entre agentes especializados coordenados por um orquestrador, introduzindo a interação entre agentes. comunicação Protocolos, agregação de resultados e resolução de conflitos como preocupações arquitetônicas adicionais.
As publicações e patentes indexadas abaixo abordam algoritmos de planejamento, arquiteturas de uso de ferramentas, sistemas de memória, protocolos de coordenação multiagente, benchmarks de avaliação de agentes e metodologias de restrição de segurança.
Esta é a nossa mais recente seleção de publicações e patentes mundiais em inglês sobre IA Agética, provenientes de diversos periódicos científicos online, classificadas e focadas em IA Agética, agente de IA, agente de IA autônomo, sistema multiagente, orquestração de agentes de IA, ciclo de planejamento de agentes, agente ReAct, planejamento baseado em cadeia de pensamento, agente de IA que utiliza ferramentas, IA com chamada de funções, memória de agentes de IA, memória de longo prazo de agentes, memória episódica de agentes, memória de trabalho de agentes, decomposição de tarefas de agentes de IA, planejamento hierárquico de agentes, autocorreção de agentes de IA, reflexão de agentes de IA, avaliação de agentes de IA, modelagem de recompensas de agentes, sandboxing de agentes de IA, salvaguardas de segurança para agentes, coordenação multiagente, protocolo de comunicação de agentes, registro de ferramentas para agentes de IA, agente executor de código, agente de navegação na web, agente com recuperação aprimorada, avaliação de benchmarks de agentes e agente com interação humana.
Self-pruning fractal computational architecture for high-performance computing on resource-constrained and noisy quantum hardware
Patent published on the 2026-07-02 in WO under Ref WO2026139942 by MARECHAL THIERRY [CA] (Marechal Thierry [ca])
Abstract: A computational architecture employing self-pruning fractal branch management for achieving supercomputer-class performance on standard hardware and noisy intermediate-scale quantum (NISQ) devices. Unlike conventional parallel computing systems requiring massive hardware resources or genetic algorithms requiring extensive population evolution, this invention utilizes hierarchical fractal doubles—modular computational units organized in self-similar tree structures—with real-time adaptive pru[...]
Our summary: The architecture employs self-pruning fractal branch management for high-performance computing on noisy quantum hardware. It achieves supercomputer-class performance with modular units organized in fractal structures, utilizing real-time adaptive pruning to eliminate non-promising branches. Applications span neural architecture search, protein folding, and quantum system modeling, achieving significant speedups while being energy-efficient and noise-tolerant.
fractal architecture, quantum computing, self-pruning, computational efficiency
Patent
Hardware-enforced agentic genai workflow orchestrator with cryptographic ethical guardrails and human-in-the-loop escalation for autonomous clinical o
Patent published on the 2026-07-02 in US under Ref US20260188502 by BICKERSTAFF III GEORGE WILLIAM [US] (Bickerstaff Iii George William [us])
Abstract: A hardware-anchored orchestration system for autonomous GenAI agents in clinical settings, implementable in ASIC or FPGA fabric to ensure deterministic enforcement independent of software execution layers. The system utilizes a hardware-isolated ethical supervisor—comprising a HSM or TPM—to monitor agentic workflows against human-configured safety thresholds stored in an ethical guardrail manifest in a silicon vault. Hardware-based logic gates detect statistically anomalous token-level entro[...]
Our summary: The system orchestrates autonomous GenAI agents in clinical environments using hardware for deterministic enforcement. It employs a hardware-isolated ethical supervisor to monitor workflows against safety thresholds. In case of a safety breach, it activates a hardwired interlock to prevent non-compliant outputs.
Hardware Orchestration, GenAI Agents, Ethical Supervision, Cryptographic Compliance
Patent
Deterministic ai agent liability firewall and insurance engine with hardware-enforced envelope binding and privacy-preserving risk pooling
Patent published on the 2026-07-02 in US under Ref US20260187731 by BICKERSTAFF III GEORGE WILLIAM [US] (Bickerstaff Iii George William [us])
Abstract: A hardware-enforced AI liability containment system binds autonomous AI agent decisions to predefined liability envelopes retrieved from TEE-sealed policy stores within trusted execution environments (TEEs) comprising Intel SGX enclaves, AMD SEV-SNP protected VMs, or ARM TrustZone secure worlds, materially altering processor states to isolate all liability computations. Real-time risk exposure is computed using trust-state signals derived from TEE-resident hardware mechanisms comprising enclave-[...]
Our summary: This system binds AI agent decisions to liability envelopes using hardware-enforced mechanisms. It computes real-time risk exposure through TEE-derived trust-state signals. Insurance pools manage excess exposure with zk-SNARK proofs, ensuring verifiable operations in secure environments.
deterministic AI, liability containment, trusted execution environments, zero-knowledge proofs
Patent
Cross-agent context management for multi-agent system
Patent published on the 2026-07-02 in US under Ref US20260186828 by MICROSOFT TECH LICENSING LLC [US] (Wang Jeffrey [us], Katarya Vivek [us], Chandla Amol [us], Ng Christopher [us], Racca David Nicolas [it], Baruch Keren [us], Bottaro Juan Pablo [es], Sachindran Santhosh [us], Mohanasundaram Gokulraj [us], Arcara Kevin [us])
Abstract: An example maps input to a first task executable by a first task agent of a multi-agent application system. A search is formulated using the input and the first task. The search is executed on a second layer of a multi-layer memory accessible to the orchestrator agent. First cross-agent context data related to the input and the first task is extracted from results of the search. The first task and the first cross-agent context data are routed to the first task agent. Output is received from the [...]
Our summary: This content describes a method for managing context in a multi-agent system. It involves mapping input to tasks executed by agents and utilizing a multi-layer memory for searches. The process includes routing data between agents to generate responses based on task execution.
context management, multi-agent system, task execution, orchestrator agent
Patent
Information gathering using an intake artificial intelligence agent
Patent published on the 2026-07-02 in US under Ref US20260187111 by MAPLEBEAR INC [US] (Vrabec Helena Ursic [us], Bernard Benjamin [us], Lei Kevin [us], Cohen Spencer Lee [us], Lee Junghoon [ca], Bowering Robert [ca], Hohenberger Taylor [us])
Abstract: [0000] A question is posed by a user of a user device as part of an online chat session with an online system. An intake artificial intelligence (AI) agent interacts with the user via the online chat session in one or more rounds of messaging to gather information that may be used by a human agent to respond to the question. At some point, the online system may identify in an output of the intake AI agent an indication that there is sufficient context regarding the question to transfer the quest[...]
Our summary: An intake AI agent interacts with users in an online chat to gather relevant information. The system assesses when enough context is available to transfer the interaction to a human agent. The human agent utilizes the gathered session information to formulate a response to the user s question.
AI agent, information gathering, online chat, human agent
Patent
Knowledge extraction system and knowledge extraction method
Patent published on the 2026-07-02 in US under Ref US20260187494 by HITACHI LTD [JP] (Suzuki Shintaro [jp], Hyodo Akihiko [jp], Sakaniwa Hidenori [jp])
Abstract: [0000] An AI agent system includes an episode memory database configured to accumulate dialogue histories with users as episode memories including a plurality of messages—, a semantic memory construction unit configured to cluster the messages included in the episode memories into a plurality of clusters and extract scenario branches based on transitions of the messages between the clusters, and a semantic memory database configured to store the scenario branches as semantic memories.[...]
Our summary: The AI agent system accumulates dialogue histories in an episode memory database. It clusters messages into groups and extracts scenario branches based on message transitions. The scenario branches are stored as semantic memories in a dedicated database.
knowledge extraction, AI agent, semantic memory, dialogue history
Patent
agentic AI and the future of electron microscopy
Published on 2026-04-10 by Vida Jamali, Amirali Aghazadeh, Josh Kacher @NATURE npj
Abstract: npj Computational Materials, Published online: 10 April 2026; doi:10.1038/s41524-026-02077-yAdvances in microscopy have long focused on improving resolution, throughput, and automation. The next transformation may lie in enabling microscopes to contribute to the reasoning that guides experiments. Recent advances in agentic artificial intelligence (AI) suggest a future in which microscopes do more than simply acquire images. Agentic systems could draw on prior knowledge, interpret experimental ou[...]
Our summary: Advances in agentic AI could enable electron microscopes to interpret data and design experiments. This transformation may shift microscopes from passive tools to active collaborators in research. The transition requires community support through open access and data sharing initiatives.
agentic AI, electron microscopy, experimental design, materials characterization
Publication
A Scoping Review
Published on 2026-02-01 by Jonathan Gibson, Praveen Chinniah, Shashank Chapala, Ojasvi Vemuri, Rajesh Botchu @MDPI
Abstract: Objectives: Artificial intelligence (AI) is a transformative development in the field of medicine. In the field of musculoskeletal radiology, agentic AI is a technology that could flourish, but currently, the limited evidence base is fragmented and sparse, and we present a scoping review of it. Methods: Parallel searches were conducted in four databases: PubMed, Embase, Scopus, and Web of Science. Search terms included all agentic AI and autonomous AI agents, as well as radiology. All papers und[...]
Our summary: This scoping review evaluates the potential of agentic AI in musculoskeletal radiology. It identifies eleven relevant studies highlighting improved decision support, workflow optimization, and image analysis. Despite promising findings, the evidence base remains limited and theoretical.
AI in Radiology, Musculoskeletal Imaging, Workflow Optimization, Decision Support
Publication











