Agentic Systems
Autonomous agents, tool use, multi-agent collaboration, and extended autonomous action. AI that can act in the world.
Key Benchmarks
Recent Papers
SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic Skill-Use
Jiayin Zhu, Kelong Mao, Yudong Guo +4 more
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
Wentao Zhang, Liliana Hotsko, Woojeong Kim +3 more
AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents
Xiangchen Cheng, Yunwei Jiang, Jianwen Sun +7 more
AgenticDataBench: A Comprehensive Benchmark for Data Agents
Zhaoyan Sun, Shan Zhong, Daizhou Wen +10 more
PACE: A Proxy for Agentic Capability Evaluation
Yueqi Song, Lintang Sutawika, Jiarui Liu +8 more
AutoTrainess: Teaching Language Models to Improve Language Models Autonomously
Zhaojian Yu, Penghao Yin, Shuzheng Gao +3 more
DuoMem: Towards Capable On-Device Memory Agents via Dual-Space Distillation
Shengguang Wu, Hao Zhu, Yuhui Zhang +2 more
AI-Model Network: Concept, Current State and Future
Li Zhetao, Zeng Xiyu, Wang Jianhui +6 more
Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning
Xuan Zhang, Zhijian Zhou, Lingfeng Qiao +6 more
ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation
Zhaoqi Wang, Zijian Zhang, Kun Zheng +4 more
Recent Milestones
OpenAI Ships GPT-5.6 and ChatGPT Work Agents
OpenAI on July 9, 2026 broadly released its GPT‑5.6 model family — Sol, Terra and Luna — after a limited government‑review preview. The company also launched ChatGPT Work, an agentic productivity tool that combines ChatGPT and Codex to automate end‑to‑end knowledge work across apps and documents. The rollout starts globally over the next 24 hours across ChatGPT and Codex.
Claude Cowork Agents Go Cloud, Web, and Mobile
On July 8, 2026, TechRadar reported that Anthropic has enabled Claude Cowork sessions to run from mobile apps and a dedicated web portal, with workflows executing in the cloud by default. Anthropic also shared usage data showing Cowork is now used more for knowledge work than for coding.
Claude Cowork agent expands to web and mobile
Anthropic began rolling out its Claude Cowork agent to web and mobile on July 8, 2026, after initially limiting it to a desktop app. The expansion lets Max-tier users assign multi‑step tasks that Cowork can continue executing even when their laptop is closed.
Proofpoint embeds GPT‑5.5 in cyber defense
Cybersecurity company Proofpoint has been selected for OpenAI’s Daybreak Cyber Partner Program, allowing it to integrate GPT‑5.5 into its managed security products and workflows. The partnership focuses on using OpenAI models for defensive tasks like threat analysis, alert triage, and incident response while keeping direct model access with trusted partners.
Microsoft $2.5B Frontier Co for AI deployments
Microsoft has created a new business unit called Frontier Company to embed AI engineering teams directly inside large customer organizations. The group will deploy over 6,000 specialists and is backed by what reports describe as a $2.5 billion internal commitment to accelerate enterprise AI projects.
Codex data proves the agent era is here
Quasa.io reports on a new OpenAI economic research paper, “The Shift to Agentic AI: Evidence from Codex,” which analyzes millions of Codex usage records. The study finds that by mid‑2026, 99.8% of internal OpenAI employee output tokens and over 60% of organisational tokens flowed through Codex agents rather than chat interfaces, with median employee token output rising 10–50x since late 2025.
Microsoft Frontier Co. Bets $2.5B on Enterprise AI
On July 2, 2026 Microsoft announced Microsoft Frontier Company, a new operating business that will invest $2.5 billion and deploy around 6,000 engineers and industry experts to embed AI systems inside large enterprises. Spanish outlet CincoDías and Chinese tech site 36Kr reported the initiative in detail on July 4, 2026.
LLM Nurse Safely Handles Cardiac Prep Calls
On July 4, 2026, npj Digital Medicine published a Mount Sinai Health System study evaluating a conversational AI assistant named Sofiya for pre‑procedure cardiac catheterization calls. Over roughly 1,600 patient calls, the customized LLM‑based agent completed about 88% of scripts successfully while reducing error rates between an initial stabilization phase and routine deployment.
GenAI.mil Hits 1.7M DoD Users, 100K Agents
The US Department of Defense said on July 2, 2026 that its internal generative AI marketplace GenAI.mil now has nearly 1.7 million users and over 100,000 custom agents. Officials plan to onboard more commercial models and deploy the service at higher classification levels under an updated "commercial‑first" procurement approach.
Microsoft Pours $2.5B, 6K Staff into Frontier Co
On July 2, 2026, Microsoft announced a new business unit called Microsoft Frontier Company backed by a $2.5 billion investment and 6,000 engineers and industry specialists. The unit will embed teams inside customer organizations to design, deploy and operate AI systems built on Microsoft and third‑party models.
Gemini Spark brings AI agents to macOS
Google’s Gemini Spark AI agent, previously available on the web and mobile, is now live in the Gemini app for macOS for Google AI Ultra subscribers in the US. A June 30 company blog post and July 1 coverage explain that Spark can automate multi‑step tasks across Mac files and apps, including remote runs initiated from a phone.
Claude Sonnet 5 brings near‑Opus agents to the masses
Anthropic released Claude Sonnet 5 on June 30, 2026 as its most agentic mid‑tier model, closing much of the gap to Opus 4.8 while remaining significantly cheaper. The model is now the default for most Claude plans and is also being rolled out via cloud partners such as Amazon Bedrock.
AWS pours $1B into agentic AI engineering
AWS announced a new Forward Deployed Engineering (FDE) organization on June 30, 2026, backed by a $1 billion investment to embed AI engineers directly inside customer teams. The group will build agentic AI systems on top of AWS services, aiming to compress deployment timelines from months to days while leaving customers self‑sufficient. ([aboutamazon.com](https://www.aboutamazon.com/news/aws/aws-1-billion-forward-deployed-ai-engineers))
Claude Sonnet 5: cheaper near‑Opus agents
Anthropic released Claude Sonnet 5 on June 30, 2026 as its most agentic mid-size Claude model, narrowing the gap with Opus 4.8 at significantly lower prices. The model is now the default for free and Pro users and launches on the Claude API and platform with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. ([anthropic.com](https://www.anthropic.com/news/claude-sonnet-5))
ARIA agent automates model experimentation loops
On June 29, 2026, CoreWeave announced ARIA, an AI research and iteration agent embedded into Weights & Biases’ Weave platform that reads experiment data, surfaces insights, and suggests model or agent improvements. ARIA enters preview alongside the general availability of Weave as CoreWeave deepens its tooling for AI developers. ([coreweave.com](https://www.coreweave.com/news/coreweave-aria-launches-as-an-ai-research-and-iteration-agent-with-autonomous-research-and-collaborative-intelligence))
Oracle Ships Agent Teams for Supply Chains
Oracle announced four new Fusion Agentic Applications for Oracle Fusion Cloud Supply Chain & Manufacturing, built around coordinated teams of AI agents, on June 29, 2026.([prnewswire.com](https://www.prnewswire.com/news-releases/oracle-adds-new-fusion-agentic-applications-to-help-customers-improve-supply-chain-performance-302812443.html)) The tools focus on inventory planning, supplier qualification, production readiness and Kanban management, and are already available to Oracle Cloud SCM customers.([zonebourse.com](https://www.zonebourse.com/actualite-bourse/oracle-annonce-quatre-nouvelles-applications-fusion-agentic-et-des-capacites-d-optimisation-des-stoc-ce7f5fded08cf62d))
GLM‑5.2: Open Chinese model hits frontier cyber
On June 28, 2026, multiple outlets reported that Zhipu AI’s open‑weight GLM‑5.2 model matches or beats Anthropic’s restricted Claude Mythos on security bug‑finding benchmarks, based on tests by security firm Semgrep. Analysts say the result, if confirmed, exposes gaps in US export controls that targeted Mythos-class capabilities while leaving competitive Chinese open models accessible.
Gemini 3.5 Flash Learns to Use Your Computer
On June 28, 2026, Sina Finance reported that Google has moved its "computer use" tool from a separate Gemini 2.5 model into Gemini 3.5 Flash as a built‑in capability, allowing the model to see screens and operate browsers, mobile apps and desktop software. The feature is in public preview and targets long‑running enterprise automation workloads, with tight integration into Google Cloud. ([finance.sina.com.cn](https://finance.sina.com.cn/stock/usstock/summary/2026-06-28/doc-iniewwyh6700122.shtml))
Gemini 3.5 Flash Gets Built‑In Computer Control
Forbes reported on June 27, 2026 that Google has integrated its "computer use" capability directly into the Gemini 3.5 Flash model, allowing the AI to see screens and control browsers, desktops and mobile apps like a human user. The upgrade, rolled out on June 24 and highlighted publicly on June 27, targets developers and enterprises building agentic workflows that automate multi‑step tasks across devices.
Sakana, 360 Ship Mythos‑Class Orchestration Alternatives
On June 27, 2026, TechCrunch reported that China’s 360 Security Technology and Japan’s Sakana AI have unveiled new AI systems positioned as alternatives to Anthropic’s Mythos and Fable 5 models, which remain restricted under a U.S. export order. 360 introduced its Tulongfeng and Yitianzhen cybersecurity tools, while Sakana AI launched its Fugu orchestration model targeting Japanese enterprises seeking frontier capabilities without U.S. export‑control risk.