The paper proposes a three‑engine architecture for reliable financial AI, combining a financial knowledge graph, a deterministic optimization engine, and a multi‑agent LLM layer. A coordination layer called context engineering selects and validates context from ERP systems, tools, and tenants to ensure numerical accuracy, traceability, and auditability. Observations from 15 deployments show that graph‑backed multi‑agent validation can reduce hallucination rates from 7‑9% to below 1%, while reusable mappings and validation policies cut deployment time from six to two weeks.
The paper investigates how large language model agents handle long‑horizon tasks and finds that verification is crucial for success but often occurs too late. It shows that sharing a common context dilutes duty‑relevant information and that simply separating duties among agents is insufficient, as agents can lose sight of the overall goal. To address these issues, the authors propose Goal‑Aware Dynamic Multi‑Agent Collaboration (GAD), which uses an execution graph to organize subtasks, keeps each agent’s context focused on its specific goal, and dynamically adjusts task dependencies and information sharing based on execution evidence, with verification and recovery checks providing feedback for repair.
Autonomous LLM agents executing multi-step workflows over personal data often misalign with user privacy expectations and leak sensitive, out-of-context information. Although Contextual Integrity (CI)...
The paper investigates how self‑evolving personal agents become more vulnerable to indirect prompt injection (IPI) attacks. Using a two‑phase methodology—first allowing the agent to learn from user feedback on a task, then testing it with a third‑party hidden instruction—the authors find that self‑evolution significantly increases attack success rates, rising by 14.0–48.8 percentage points with skill evolution and 3.0–39.3 percentage points with reinforcement learning. The study demonstrates that the very mechanisms designed to improve agent performance also degrade their robustness to malicious manipulation.
PCM is a protocol that attaches verifiable evidence to every claim sent between AI agents, using immutable Merkle logs and signed attestations that the sender cannot tamper with. Receivers evaluate messages offline and classify each claim as proven, unproven, or failed, ensuring that self‑reported claims can never be proven and that authority is traceable through recorded human approvals. In a benchmark of 926 real messages, PCM flagged only 10 stale‑evidence lies (1.4%) while other protocols accepted many more, and verification runs in about 1 ms with perfect agreement between independent verifiers.
GitHub Octoverse 2026 reports a dramatic rise in activity, with merged pull requests tripling to 220 million per month, commits more than tripling to 7.38 billion per month, and new repositories doubling to 30.3 million per month. The report also provides industry‑specific data and highlights repositories focused on interoperability or reverse engineering. Methodology details include country‑level push adjustments using LLM‑derived ISO codes and NAICS sector classification via a multilingual embedding model.
The article discusses how to integrate AI into complete business processes, covering topics such as simplifying workflows, linking knowledge and rules, coordinating multiple AI agents, and achieving measurable business outcomes.
By Weiwei Hu
Simon Willison discusses a new inflection point in October 2026 for creating video games using AI, noting that Claude Opus 5.5 now reliably produces compelling 3D models, music, and especially game design. He reports having successfully "vibe coded" four enjoyable games without inspecting the code, and highlights that the real challenge is now generating novel, delightful concepts that AI alone could not conceive. The author compares Opus 5.5’s fun factor to GPT‑6 Astra, which he finds lacking in that regard.
OpenAI and ServiceNow have broadened their partnership, integrating OpenAI models and agents into ServiceNow and incorporating ServiceNow workflows into ChatGPT and Codex. This expansion enables deeper collaboration between the two platforms, allowing AI-driven capabilities to be embedded within ServiceNow’s enterprise workflow environment and vice versa.
The study compares deterministic static application security testing (SAST) with AI coding agents and AI security harnesses using the VulnBench 2.0 benchmark of 108 vulnerabilities across multiple languages. Nine productized configurations were evaluated, with Snyk Code achieving 82.4% macro recall and Claude Code at 79.3%; together they covered 101 of 108 vulnerabilities, exceeding any single configuration. The results show that AI engines’ effectiveness depends heavily on their harness policies and that AI tools can uncover real vulnerabilities absent from the reference set.
The article discusses how engineers should view the lack of benefit from current AI tools as a finding rather than a failure. It encourages documenting these observations, sharing them with peers and managers, and actively seeking solutions. By explaining where AI falls short, engineers demonstrate the judgment expected of them.
Legora is developing an agentic operating system for legal work that leverages OpenAI models, ChatGPT, and Codex. The platform accelerates legal workflows, expands contract analysis capabilities, and enables lawyers to concentrate on higher‑value tasks. By integrating these AI tools, Legora aims to streamline and enhance the efficiency of legal processes.
The paper investigates how fabricated claims spread, persist, and can be mitigated in multi‑agent social networks, using a controlled framework that simulates agent communication, private memory, and a shared web. Experiments with 540 fabricated claims across eight domains show that false beliefs can reach up to 96% of agents, remain stable even after the original fabricators withdraw, and are more effectively countered by targeted corrective interventions than by preventive measures. The study highlights the safety challenge posed by persistent false beliefs in such networks and the need to address both their initial dissemination and the evidence that sustains them.
By Dianzhi Yu, Hongru Wang, Ziheng Xiong, Yanyu Chen, Zixing Song, Jianting Chen, Zhenfei Yin, Jeff Z. Pan, Amos Storkey, Irwin King
Organon is a recursively self‑improving agent designed for scientific discovery that autonomously revises its search strategies, population management, diagnostics, memory, and worker coordination during execution. By diagnosing and correcting its own adaptation procedures, the agent continuously refines the mechanisms guiding its future self‑improvement. Across 21 computational discovery tasks—including algorithm design, GPU kernel optimization, and systems optimization—Organon outperforms competing frameworks, achieving a 20.5% mean performance gain on ALE‑Bench Lite and up to 1.67× speedups on KernelBench.
By Cheikh Ahmed, Zhiyu Zhang, Kun Mao, Yong Zhang, Zirui Zhou
FeatureLoop is a benchmark that evaluates coding agents on 61 human‑reviewed tasks drawn from 34 repositories in five programming languages. Each task tracks a development goal through iterative feedback—adding or clarifying requirements and reporting failures—while cumulative executable checks monitor new behavior, repairs, and preservation of earlier functionality. In tests of 12 state‑of‑the‑art models, the best agent completed 72.1% of tasks with both feedback types, but only 15.4% of failures were recovered within two revisions, highlighting issues such as incomplete repairs, regressions, and token inefficiency.
By Jia Feng, Alyssa Wu, Cuiyun Gao, Yunlong Wang, Songyuanyi Lu, Zhichao Hu
The paper introduces ConCal, a recursive continuation-value estimator designed to improve action-value calibration in long-horizon reinforcement learning. ConCal matches visits by state-action-state transitions, merges shared continuation evidence with individual transition estimates, and propagates calibrated values backward to refine earlier action-value estimates. Experiments show ConCal reduces local credit distortions, lowers error in value differences, and improves success rates over GiGPO on ALFWorld and WebShop benchmarks.
By Yingjian Chen, Lang Zhou, Zixin Zhao, Pingfan Li, Weishi Zheng, Zhilin Zhao
The paper presents a framework called SNPE‑TTT for adapting pre‑trained neural posterior estimators, specifically conditional normalizing flows, to new inference tasks in agent‑based models (ABMs). It explores four adaptation strategies—full‑parameter adaptation, low‑rank adaptation, and two gradient‑based update spaces—and evaluates them on economic simulation models with both tractable and intractable likelihoods. Results show that the framework can improve posterior recovery with a limited number of target simulations, though performance varies by task and adaptation method.
By Junyao Zhang, Jinglai Li, Junqi Tang
The paper introduces a continuous‑time actor‑critic framework for jump‑diffusion control problems, using a time‑inhomogeneous q‑function and occupation measure to derive a policy‑gradient representation that handles time‑dependent drift, volatility, and jump terms. The actor is modeled with conditional normalizing flows, allowing flexible non‑Gaussian continuous‑action policies while maintaining exact likelihoods for entropy regularization. Experiments on linear‑quadratic control, Merton portfolio optimization, and a multi‑agent portfolio game show stable learning, accurate policy approximation, and good scaling with dimension and agent count.
By Liya Guo, Ruimeng Hu, Xu Yang, Yi Zhu
The paper presents a Staged Multi-Agent Training (SMAT) method that jointly trains a musculoskeletal human model and a hip-exoskeleton controller in simulation to learn coordinated assistance strategies accounting for human-device co-adaptation. The resulting policy was transferred to a physical hip exoskeleton and tested on eight healthy adults without subject-specific retraining, showing a 19.7% reduction in net metabolic rate compared to a passive device and a positive-power ratio of 0.984 across walking speeds and terrains. These findings demonstrate that simulation-based multi-agent co-adaptation can yield a single transferable exoskeleton policy that provides meaningful physiological benefits in real users.
By Yifei Yuan, Jakob Wolf, Ghaith Androwis, Xianlian Zhou
FUSION is a skill‑based research agent that automates the use of publicly available nuclear‑physics codes by retrieving the code, running verified inputs, parsing outputs, and ensuring results meet specified benchmarks. The current release supports twenty different codes across various nuclear physics domains and includes a searchable collection of literature references. Documentation and source code are freely available under an MIT license.
By Jin Lei