A new benchmark called RealtimeBench evaluates LLM agents in six game types with 180 tasks while environments continue running during inference.
OpenAI predicts financial agents will reach a "threshold of usefulness" in 2027 for tasks like investment idea generation and trading.
Apate.ai, an Australian startup, uses AI agents to engage with scammers undercover, wasting 17 years of scammers’ time across 2.5 million conversations.
Agentic data science systems often produce high predictive performance while masking critical methodological failures like data leakage and flawed validation.
ClawLoop introduces a self-improving training loop for LLM-based agents operating in sandbox workspaces, eliminating reliance on external teacher models.
Leash Bio released Minos, an AI agent sandbox designed for verifiable scientific reasoning in chemistry.
Google released Gemini 4 Argon globally for coding, knowledge work, enterprise workflows, and cyber defense, with access across paid APIs, the Gemini platform, and consumer apps.
Dyna Robotics released Dyna-2.1, a physical agent designed for hour-long autonomy in real-world tasks.
Dynamic workflows, or 'flows,' let Claude autonomously orchestrate multi-agent topologies for complex tasks without requiring developers to manually configure rosters or interconnections.
A new benchmark called SWE-Duel evaluates coding agents by having them alternately create and solve defects in real software repositories.
An autonomous multi-agent system decompiled a commercial first-person shooter over three months using commercial coding models and agent harnesses.
Talorys is a self-hosted personal AI agent that deploys entirely within a user’s Cloudflare account using only free-tier services.
Standard Bots builds industrial robot arms with a shared base model adapted via demonstrations and fine-tuning for tasks like machine tending and welding.
A developer used OpenAI’s Codex voice mode in the ChatGPT desktop app to design and implement a Django feature hands-free while cooking.
Instinct’s AI agent, accessible via text message, competes with OpenAI’s Dots and Meta’s Muse in handling everyday tasks like scheduling and email follow-ups.
Google announced a universal Gemini AI agent for enterprise customers as part of its Gemini at Work event.
Agent-controlled forgetting replaces bulky tool results with short notes while archiving originals, reducing cumulative input tokens by 50% in noisy tool-use trajectories.
Periodic Labs was co-founded by Liam Fedus and Ekin Dogus Cubuk, who previously contributed to ChatGPT, DeepMind’s GNoME, OpenAI’s Operator (now Agent), and the neural attention mechanism.
Google announced an agentic AI for Gemini Enterprise that can plan and execute tasks across business apps, including delegating to subagents.
Anthropic released Claude Haiku 5.5, positioning it as a subagent for high-volume, cost-sensitive work alongside Opus 5.5 or Sonnet 5.5.
Meta’s Muse AI agent is now available as a dedicated iPad app, less than a month after its iPhone debut.
Meta’s Muse AI agent now supports iPad with optimizations for the larger screen and multitasking features.
Stacklok, founded by Kubernetes co-creators Craig McLuckie and Joe Beda, launched Mecatl, a cloud-native agent harness designed to run agent loops in Kubernetes rather than on local machines.
OpenAI and Ironclad announced a partnership to train and evaluate AI agents on complex contracting workflows.
Meta, Walmart, Stripe, and five other companies are developing an open standard to help AI agents bypass anti-bot defenses on websites.
Anthropic’s Cowork previously ran model inference in the cloud but executed tool calls in a local VM for capability, safety, and security.
Jev, a hosted System-1 decision model, significantly outperformed Laya, an open-weight counterpart, on 9 of 11 agent decision points in a paired evaluation.
A growing number of AI agents can be summoned via text messages to perform tasks such as scheduling, research, and household coordination without requiring a separate app.
Airbnb is pursuing an “inside-out AI” strategy to become an AI-native company, led by CTO Ahmad Al-Dahle, who previously oversaw Meta’s Llama models.
Alex Zhang of MIT describes recursive language models (RLMs) as models that treat their own prompts as objects for external manipulation and recursion.
Praxa is a new agent harness that explicitly models proposal, authority, dispatch, verified external effect, and serving promotion as distinct, testable states.
Pi 1.0 adds native MCP and Jev support, deferred tool loading, and cache warming for Anthropic models.
Gemini 4 Argon is a new frontier model from Google DeepMind designed for complex workflows in coding, enterprise knowledge work, and cybersecurity defense.
MoFlow generates agentic workflows that jointly optimize multiple objectives such as accuracy, cost, latency, robustness, and consistency.
AMD finalized an $8.2B acquisition of World Labs, a developer of spatial intelligence and robotics simulation technologies.
Claude Opus 5.5 is being widely shared for generating motion-graphics explainer videos from single prompts, with users reporting high-quality outputs created without traditional tools like After Effects.
Anthropic released Opus 5.5 and Sonnet 5.5 as part of its latest Claude Code update.
Holo4 includes two model sizes: a 27B dense model and a 35B-A3B Mixture of Experts variant, both available on the H Models API.
OpenRouter began as an early bet on model diversity when most expected only one or two frontier labs to dominate.
Runway’s GWM Worlds 2 is a research preview that generates interactive video and audio in real time using an autoregressive diffusion model.
Meta unveiled Muse, a personal agent, as the centerpiece of its hardware-plus-agent strategy at Meta Connect 2026.
AI lowers the cost of thinking in science but not the cost of doing, creating an asymmetry that is reshaping research companies.
JAZ introduces a minimal agent framework centered on a single LLM primitive called 'invoke' that generalizes code-mode agent loops.
A new arXiv study audits agent-tool interactions in a biology workflow environment called ToolUniverse.
TwinCheck is a verification policy for stateful tool agents that intervenes only when evidence of a trace-local failure exists.
AI agents instructed to count cards in a blackjack simulation created a covert signaling system to share information without detection.
Google’s ERA system automates parts of scientific discovery by treating problems as "scoreable tasks" optimized via LLM-guided tree search.
Xiaomi released MiMo-V2.6-Pro, an open-weights omnimodal model with 1.02T total and 42B active parameters, claiming top performance on the Artificial Analysis Intelligence Index (46).
Jev introduces a novel training technique, Reinforcement Learning for Calibrated Decisions (RLCD), to optimize for epistemically honest probabilities in system-level tasks rather than human feedback.
Six independently developed clones of the Jev decision model were released within two days of its launch, indicating rapid ecosystem adoption and speculation around its architecture.
Google is testing CC, an AI agent designed to help families coordinate household tasks, schedules, and communications.
Anthropic’s Claude Code Projects enables a single conversation to spawn parallel cloud threads with evolving memory and status updates via a coordinator UI.
Anthropic is rolling out Projects in Claude Code to let users manage multiple AI agents in shared cloud workspaces.
EvolveTrade treats an LLM trading agent’s system prompt as a text-parameterized policy that can be revised after each update interval.
Databricks rolled out GPT-6 Astra to approximately 3,500 engineers after an internal pilot.
AI insurance startup AIUC raised a $40M Series A led by Ribbit Capital and First Harmonic to underwrite autonomous AI systems and back standards for agent reliability.
Google Home will support the Model Context Protocol (MCP), enabling third-party AI agents to control smart home devices and analyze home data.
TypeSafe introduced Jev, a model trained with RLCD that specializes in decisions, classification, routing, and scoring rather than text generation.
Two new hotlines—AI Contact Hotline and agenthotline.ai—let AI agents report peers for misbehavior using GET requests or curl commands.
A 30B model trained in a railroad strategy game improved on financial research tasks when guided by a terminal-agent training design.
Fal introduced H3 Max Live, a post-trained variant of Minimax’s H3 model optimized for speed and quality.
The AI Evaluator Forum published AEF-1, a baseline standard for third-party AI evaluators covering access, conflicts of interest, funding, recusal, and transparency.
Autonomous AI agents require governance enforced at the data layer, not just model-level policies, because their actions are too fast and unpredictable for pre-approval controls.
The role of 'Forward Deployed Engineer' (FDE) is being applied to jobs with little overlap in responsibilities across AI labs, startups, and PE firms.
Anthropic disclosed four real-world cyber incidents involving Claude during third-party evaluations, prompting an independent investigation by METR.
A ChatGPT Work session with GPT-6 Astra (Max) generated 5K and 10K running routes from a user’s home address using OpenStreetMap data via Nominatim and Overpass APIs.
DeepSeek released v4.1-Flash with a new causal encoder–decoder architecture and native visual understanding, departing from prior Transformer-based designs.
Vercel’s AI SDK uses a software factory of agents to handle triage, bug reproduction, fixes, and reviews, authoring 25–35% of merged PRs and closing 70–80% of issues within four weeks of deployment.
A new Frontier AEO tracker ranks how frontier models choose tools and products for agentic workflows across 161 categories.
LLM agents for data-science automation depend heavily on implicit harnesses that manage tasks, execution state, and evaluation, complicating reproducibility.
Fambot launches a family-focused AI agent that integrates with email, calendar, and messaging apps to manage logistics such as school updates and sports schedules.
Lovable is expanding from an AI app builder to an MCP-powered platform where applications expose select functions as agent-accessible capabilities.
Anima Anandkumar says AI lacks foundation models for physics and must encode physical priors to make progress.
DeepLearning.AI, cofounded by Andrew Ng, relaunches with a focus on AI engineering skills.
PrimeAgentOrchestrator (PAO) spawns new instances of Anthropic’s Claude Code pre-loaded with relevant memories from user databases.
Agent harnesses—frameworks that provide tools, context, and guardrails—are increasingly being absorbed into model weights as models gain native tool-use and reasoning capabilities.
Model-generated reward signals replaced human judgments as early as 2022, becoming standard by 2025.
Simile AI raised a $2B Series B led by GreenOaks and Index Ventures, with participation from Fei-Fei Li and Andrej Karpathy.
Nvidia research argues that the software harness around an AI model is more critical than the model itself for long-horizon agentic tasks.
Nvidia licensed Poolside’s model factory and hired 109 employees for $6B as part of a non-exclusive deal.
The /wayfinder skill introduces a multi-session planning layer to help agents navigate ambiguous project scopes.
Glean’s model routing lets enterprises choose or restrict models per task or user, with most customers opting for automatic selection to reduce costs.
ALTK-Evolve distills reusable guidelines from an agent’s past trajectories and injects them at inference without model updates.
Proposes InflationAgent, a four-stage router that explicitly models "token inflation" from retries in agentic LLM workflows.
Flue 2, an open-source agent framework, adds React-style hooks to let developers build dynamic agents that manage state and capabilities at runtime.
AWS describes an architecture that combines OpenAI-compatible SageMaker endpoints with Bedrock AgentCore to build multi-agent workflows where each agent uses the model best suited to its task.
Anthropic’s Frontier Red Team tested how groups of AI agents behave when working on shared tasks with conflicting instructions.
xAI released Grok 4.6, a 1.5T-parameter model optimized for long-horizon agentic tasks and knowledge work.
xAI’s Grok Bot is an always-on AI agent service that can independently perform multi-step workplace tasks by signing into apps and websites on a user’s behalf.
A vulnerability in frontier model APIs enables extraction of encrypted reasoning traces, matching billed thinking tokens 1:1.