AI News

Google’s Gemini 3.8 Flash improves agent and software engineering performance while retaining extremely fast output. It scores 59 on Artificial Analysis’s Intelligence Index, ties several larger frontier models, and reaches 74% on DeepSWE v1.1. More reasoning and tool calls raise its average benchmark task cost by about 40% to $0.58, but the model remains highly cost-efficient for its capability level. Google also introduced Gemini 3.8 Flash Cyber for vulnerability discovery and patching, with restricted initial access.

A benchmark study finds that persistent memory can make AI agents over-trust stale information. Larger Qwen3 models were especially vulnerable when old notes appeared recent and conflicted with authoritative tools. Metadata helped capable models, while smaller models often needed conflicts resolved in advance. The pattern also appeared in Llama-Instruct models and external datasets.

ToolGate is an executable pipeline for building scientific benchmarks that require specialist software. It validates generated items with reproducible scripts, screens out questions solvable without tools, and tests survivors with a tool-using agent. In a FEniCSx experiment, 500 attempts produced 128 unique protocol survivors after GPT-5.5 screening and Codex-based solving. The process makes benchmark acceptance auditable while preserving expert control over domain design and final review.

Researchers introduce Reliable Enterprise Agent Deployment (READY), a framework for qualifying AI agents for enterprise workflows. It evaluates candidate oversight policies, selects the minimum-cost policy that meets a reliability target, and statistically validates the result on held-out cases. In a clinical-audit study of 16 agent systems and 750 cases, systems with autonomous accuracy of 72.8% and 72.5% required 39.2% and 29.6% human review, respectively, to reach the same 76% reliability target. READY reframes evaluation around the conditions and cost of dependable deployment.

Researchers propose Belief-Calibrated Optimization, which stores and revises an LLM optimizer’s beliefs about how the environment responds to edits. The approach improves training pass rates across five agent benchmarks and often transfers better after the target model is swapped. An ablation indicates that the document’s information content drives the benefit.

Anthropic is under pressure to add native AGENTS.md support to Claude Code after Shopify CEO Tobi Lütke threatened an internal ban. AGENTS.md is used by more than 60,000 open-source projects and multiple coding tools, while Claude Code relies on CLAUDE.md. Teams can bridge the formats with imports, symlinks, and automation, but those workarounds add maintenance costs in large monorepos. Anthropic says it is improving customization, without clearly promising native recursive support.

The paper introduces invalidation contracts for LLM agents that cache API-error recovery suggestions across episodes. Version stamps and cacheability hints let clients remove stale entries after server-side drift while preserving valid ones. Evaluation across seven models and about 9,400 episodes found that row-level invalidation improved compliance by up to 66.7 percentage points and recovered 29–33% of baseline token cost on four models. The protocol added 15% to response payloads and produced deterministic validity with no contract failures.

Researchers present FAIRY, a full-stack agentic system deployed at a soybean research farm. It coordinates the complete production cycle, integrates machinery, sensors, drones, satellites, weather data, crop models, and farm records, and represents changes as events in a shared spatiotemporal process engine. Nine agent controllers are evaluated across 100 full-season scenarios, with metrics covering operational success, path correctness, token cost, and edge-device runtime.

OpenAgentFlow provides a shared control-plane and action-plane architecture for governing heterogeneous AI agent actions. It normalizes multiple action types into an AgentEvent stream and applies policies before execution, while preserving provenance and audit data. Its Android evaluation reports 94.0% accuracy, a 95.3% attack-block rate, and a 92.9% trace-adjusted pass rate in traced cases.

AI is moving into physical environments as robotics models become more reusable and hardware becomes cheaper. Yunzheng Li argues that long-term value will shift from robot bodies to intelligent models, real-world data, middleware, tools, and developer ecosystems. He also describes Agent-native development, where AI agents configure environments and run robotics demos. China’s manufacturing strengths and the United States’ model and software ecosystems may provide complementary foundations for global robotics platforms.

Coinbase CEO Brian Armstrong says AI agents can now trade equities on the platform, in addition to crypto and derivatives. Coinbase documentation says equity features are beginning to roll out for Agents through the Coinbase MCP, enabling stock research, sentiment analysis, performance tracking, and trade execution. The announcement marks a move toward “agentic finance,” but the supplied evidence does not independently confirm availability, geographic scope, trading limits, safeguards, or the full operational scope.

Tencent expanded the Hy4preview inference cluster after the model’s WorkBuddy debut triggered heavy demand and queueing. The company also extended limited-time free access to Hy3 and other existing models through September 30, 2026, while warning that peak-time queues may recur.

ReToolSQL combines supervised training on verified reasoning traces with agentic reinforcement fine-tuning over multi-turn tool-use trajectories. On BIRD-SQL, its Gemma 4 31B model achieves 74.32% single-pass execution accuracy and 74.77% with self-consistency.

Researchers introduce GMA, a benchmark for evaluating general mobile assistants in realistic, challenging scenarios. It spans seven open-source applications and 300 tasks across four difficulty tiers. Tests of eight frontier models show sharply declining performance as workflows become more complex. Ablation studies find that context retention and explicit state tracking can improve results, although effective harness designs vary by foundation model.