AI Agent News
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Latest industry news
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Key events timeline
OpenClaw erupts on GitHub
OpenClaw hits global GitHub Top 10 in 10 days, outpacing the Linux kernel star growth
Meta acquires Manus for $2B
Meta acquires Manus AI for $2B, locking in the general-purpose Agent race
DeepSeek-V3 open-sourced
The value king, at just 5% of GPT-4 cost
Manus goes viral overnight
The world's first general-purpose AI Agent draws unprecedented attention
OpenAI Deep Research
OpenAI ships a deep-research Agent that generates professional reports in one click
MCP Servers pass 500
The MCP ecosystem erupts — 500+ servers built in 3 months
DeepSeek-R1 stuns the world
Open-source reasoning model at just 3% of OpenAI cost, reshaping the global AI landscape
MCP protocol born
Anthropic releases the Model Context Protocol, the de facto standard for Agent interfaces
Claude Computer Use
Anthropic lets AI directly control the computer screen for the first time, opening a new paradigm
Replit Agent full-stack automation
Natural language to a shipped product, aimed at non-engineers
Cursor ARR passes $100M
The fastest-growing SaaS ever, the new king of AI coding tools
Claude 3.5 tops SWE-bench
The strongest coding AI, bug-fixing at a junior engineer level
Devin launches
The world's first autonomous AI software engineer, able to complete full coding tasks on its own
Xiaomi Releases HarnessX: A Composable, Adaptive, and Evolvable Runtime Framework for AI Agents
The Xiaomi Darwin Agent Team published a paper on arXiv in June 2026, proposing HarnessX—a composable, adaptive, and evolvable runtime framework (Harness) generation system for AI agents. HarnessX addresses the pain points of traditional agent frameworks such as manual construction, coupled architecture, and inability to co-optimize with models. It achieves automated framework optimization through three core modules: - **Componentized Composition Architecture**: Abstracts the Harness into standardized processors and eight lifecycle hooks, dividing into nine optimization dimensions (model selection, context assembly, memory management, tool ecosystem, execution environment, evaluation and reward, control and security, observability, training bridging), enabling free plug-and-play and safe reuse of components via type constraints. - **AEGIS Adaptive Evolution Engine**: Maps framework evolution to symbolic-space reinforcement learning, using a four-stage pipeline of parsing, planning, evolution, and evaluation to specifically address reward hacking, catastrophic forgetting, and insufficient exploration, with variant isolation strategies for heterogeneous task scenarios. - **Framework-Model Co-Evolution Mechanism**: Builds a shared trajectory buffer and uses cross-framework GRPO algorithm to simultaneously optimize the framework and model, breaking through the capability ceiling of single optimization paths. Experiments were conducted on five mainstream agent benchmarks (GAIA, ALFWorld, WebShop, τ³-Bench, SWE-bench Verified) with three large models of varying capabilities (Claude Sonnet 4.6, GPT-5.4, Qwen3.5-9B). Results show: - HarnessX brings an average performance improvement of **14.5%**, up to **44.0%**, with weaker model baselines benefiting more significantly; - Framework-model co-evolution yields an additional **4.7%** gain; - Variant isolation effectively prevents performance degradation and reduces token consumption by approximately **18%**. This research demonstrates that optimizing the runtime framework is an efficient path to enhance agent capabilities beyond model scaling, with the entire solution balancing automation, stability, and auditability. The paper will open-source the complete code, providing a new paradigm for self-evolving agent engineering.
Tsinghua Team Proposes OpenRath: A Session-Centric Multi-Agent Runtime Framework
The Rath Team from Tsinghua University, Sun Yat-sen University, and The Chinese University of Hong Kong has open-sourced OpenRath, a multi-agent runtime framework that treats Session as a first-class citizen instead of Agent. This design addresses issues like chaotic state management and debugging difficulties in multi-agent systems. Inspired by PyTorch, OpenRath maps Session to Tensor, Agent to Layer, and Sandbox/Memory to Device/Parameter, enabling modular and composable agent clusters. The framework supports Multi-Agent Multi-Session (MAMS) mode, with capabilities for session forking, merging, and reuse, and includes built-in tools for file, shell, and code execution. OpenRath v1.2.1 is available on PyPI under the BSD-3-Clause license. ## Background: State Management Challenges in Multi-Agent Systems As the number of agents increases, traditional agent-centric designs expose serious issues: each agent maintains its own context, task forks become hard to track, and states like tool calls, memory, and sandboxes are scattered. When the system scales to dozens or hundreds of agents, debugging, reproduction, and orchestration become extremely difficult. Existing frameworks like AutoGen, CrewAI, and LangGraph mainly address inter-agent communication but do not deeply handle session state persistence and routing. ## Core Design: Session as a First-Class Citizen OpenRath proposes using Session as the carrier of state and evidence, rather than simple chat history. A Session records the complete execution trace, including tool call parameters and results, file changes, sandbox identity, failure paths, approval signals, etc. Agents are designed as transformation layers on Sessions, following the interface `forward(session) -> session`, allowing them to be stacked and nested like neural network layers. Workflow corresponds to PyTorch's Module, enabling chaining multiple agents, forking sessions, compressing context, etc. ## Key Features: MAMS and Pluggable Backends - **Multi-Agent Multi-Session (MAMS)**: Supports a four-quadrant pattern of Agent count × Session count, where the real need for forking, merging, and reusing lies in Session data flows. - **Pluggable Backends**: Sandbox (e.g., sandbox environment) and Memory (e.g., long-term memory) are replaceable backends decoupled from computation logic, similar to PyTorch's device concept. - **Tool Abstraction**: FlowToolCall unifies tool definition and execution, with built-in file, shell, and code execution tools, and supports MCP protocol adaptation. ## Comparison with Existing Frameworks | Framework | Core Abstraction | State Management | Use Case | |-----------|------------------|------------------|----------| | AutoGen | Multi-agent dialogue | Message list | Chat-style collaboration | | LangGraph | Graph state + Supervisor | Graph state | Routing and control | | OpenRath | Session + Workflow | Persistent Session | Production-grade agent clusters | OpenRath emphasizes that agent clusters are not group chats but a runtime control plane built on persistent Session states. ## Open Source and Ecosystem - **Version**: v1.2.1 (PyPI) - **Installation**: `pip install openrath` - **License**: BSD-3-Clause - **Resources**: Website (openrath.com), Docs (docs.openrath.com), Blog (blog.openrath.com), GitHub (github.com/Rath-Team/OpenRath)
Tencent Hunyuan Open-Sources UniRL: A Unified Multimodal Reinforcement Learning Training Framework
On June 17, 2025, the Tencent Hunyuan team officially open-sourced UniRL, a distributed reinforcement learning (RL) post-training framework for multimodal generative models. Led by Pang Tianyu's team, UniRL aims to address the fragmentation in current AIGC multimodal RL training—where image diffusion models, video generation, VLMs, and LLMs each have their own independent tech stacks, leading to engineering duplication and hindered algorithmic innovation. ## Background: The "Silo Dilemma" of Multimodal RL With the rapid development of models such as Stable Diffusion, FLUX, Wan, and HunyuanVideo, the capabilities of AIGC are constantly expanding, but the RL training infrastructure lags significantly. Compared to LLM RL training, multimodal generative RL faces four major challenges: - **Fundamentally different generation processes**: LLMs handle discrete tokens, while image/video generation involves continuous latent space denoising trajectories; unified multimodal models mix token generation with latent denoising, making credit assignment and policy updates more complex. - **Unstable system loop**: Rollout, log-prob replay, and policy updates span multiple models and backends. The training side must strictly replicate the sampling-side conditions; otherwise, Training-Inference Mismatch occurs, introducing policy gradient bias. - **Heavier reward system**: Multimodal RL rewards rely on multimodal evaluation chains such as VLM, OCR, aesthetic models, and video understanding models, which are costly. - **High trajectory storage and memory pressure**: Intermediate products are high-dimensional latents, noise, timesteps, etc., which scale rapidly with resolution, frame count, and denoising steps in video generation. These challenges have led to the industry practice of "one model, one training code," with developers spending significant time on repetitive engineering. ## Core Design of UniRL: Unified Abstraction and Reusable Skeleton UniRL is not tied to a single model family, algorithm, or training stack. It is built around Ray worker groups, Hydra flat recipes, composable training backends, and pluggable rollout engines, abstracting the multimodal RL closed-loop contract: rollout → reward → advantage → train → weight-sync. The framework uses typed rollout data models (tracks) to represent generation trajectories at different stages: AR stages use TextSegment, image generation stages use LatentSegment. Different tracks are connected via parent-child relationships, naturally supporting chain processes like Bagel and HunyuanImage 3.0, where AR text reasoning precedes DiT image generation. ## Supported Models and Algorithms UniRL covers mainstream multimodal generative models: - **Image generation**: SD3/3.5, Qwen-Image, Z-Image, FLUX.2-Klein - **Video generation**: HunyuanVideo 1.0&1.5, WAN series - **Large language models**: Qwen3 series - **Multimodal understanding models**: Qwen-VL series - **Native unified multimodal models**: HunyuanImage 3.0, Bagel - **Composable models**: LLM/VLM + Diffusion Prompt-Enhancer architecture Built-in RL algorithms include: - Policy-gradient family: FlowGRPO, DanceGRPO, MixGRPO, LLM/VLM GRPO - Forward-process family: DiffusionNFT - Tencent Hunyuan proprietary algorithms: Flow-DPPO (for flow/diffusion models, using stepwise KL divergence proximal constraints instead of PPO ratio clipping) and DRPO (using advantage-weighted smooth policy shift regularization instead of hard clipping/masking) Reward components integrate CLIPScore, GOT-OCR-2.0, PickScore, HPSv2/v3, ImageReward, UnifiedReward, GenEval2, WISE, VideoPickScore, VideoAlign, etc. ## Impact and Significance UniRL pushes the repetitive, error-prone, and hard-to-reuse system engineering problems in multimodal RL training down to the framework level, allowing developers to avoid repeatedly rewriting rollout, reward, trajectory transmission, and training alignment logic. The framework is still under active iteration, with plans to improve the core training loop, expand rollout engine support, and optimize large-scale training performance.
MCP Releases Authorization Spec 1.1: Enabling Enterprise-Grade Secure Agent Deployment
The MCP Working Group has officially released Authorization Spec 1.1, introducing OAuth 2.0 integration, fine-grained permission control (per-tool scopes), audit logging, and token revocation mechanisms. This update addresses the biggest concern enterprises have when deploying agents—'AI permission runaway.' Now administrators can precisely control which tools each agent can invoke and which data it can access, with complete operation records. Salesforce, Atlassian, and GitHub have announced that their MCP Servers will upgrade to the new specification, and analysts predict a surge in enterprise agent procurement in Q3 2025.
Google Releases ADK: Empowering Any Developer to Build Multi-Agent Systems
Google has officially open-sourced the Agent Development Kit (ADK), a Python framework designed for building, testing, and deploying multi-agent workflows. ADK natively supports Gemini models, with built-in tool calling, state management, inter-agent communication, and deep integration with Google Cloud Vertex AI for one-click production deployment. Unlike LangChain's general-purpose design, ADK's agent orchestration centers on a 'conversation tree,' modeling real business processes more closely. Within the first week of release, it garnered over 8,000 GitHub stars.
MCP Ecosystem Explodes: Over 500 Servers Released by Community
Just 3 months after the release of the MCP protocol, the number of community-contributed MCP Servers has exceeded 500, covering databases, browsers, code tools, SaaS platforms, and more. Major AI editors like Cursor and Windsurf have announced support for MCP, and Claude Desktop has become the most popular MCP client with over 2 million users.
Dify Open Source LLMOps Platform Surpasses 50,000 GitHub Stars
The open-source AI application development platform Dify has exceeded 50,000 stars on GitHub, becoming one of the most popular AI application building tools. Dify supports visual workflow orchestration, RAG knowledge bases, agent construction, and provides comprehensive LLMOps capabilities. Over 100,000 AI applications have been built based on Dify.
MCP (Model Context Protocol) Officially Released, Becoming the New Standard for Agent Interfaces
Anthropic released the open protocol MCP (Model Context Protocol), aiming to standardize the interaction between AI models and external tools/data sources. It quickly gained support from major vendors like OpenAI, Google, and Microsoft, becoming the de facto standard for the AI Agent ecosystem, with over 500 MCP Server implementations.
CrewAI Raises $18M as Multi-Agent Collaboration Framework Rises
AI development framework company CrewAI announced the completion of an $18 million Series A funding round. Its multi-agent collaboration framework has surpassed 20,000 GitHub stars in the past year and is used by thousands of enterprises to build automated workflows. CrewAI enables multiple AI agents to collaborate like a team through role-based division models, addressing the capability bottlenecks of single agents.