Harness engineering is emerging as a key discipline in AI agent development, emphasizing components like filesystems, memory, and retries beyond just models. OpenAI's Codex is expanding agentic coding workflows beyond software engineering, including codebase understanding and bug triage. Tooling trends show convergence on multi-agent orchestration, observability, and remote control, with GitHub Copilot, Cursor, and LangChain advancing these capabilities. The Hermes Agent v0.9.0 release introduces a local web dashboard and enhanced security, gaining community traction over OpenClaw for UX and efficiency. The open agent ecosystem is growing with projects like Open Agents and DeepAgent providing modular stacks and runtimes.

GLM-5.1 has reached #3 on Code Arena, surpassing Gemini 3.1 and GPT-5.4, and matching Claude Sonnet 4.6 in coding performance. Z.ai now holds the #1 open model rank close to the top overall. The advisor pattern, combining a cheap executor with an expensive advisor, is gaining traction, improving performance and efficiency in models like Haiku + Opus and Sonnet + Opus. Alibaba's Qwen Code v0.14.x introduces orchestration features including remote control channels, cron tasks, and sub-agent model selection. Model routing is becoming a product-level concern due to specialization and spikiness in top models such as Opus and GPT-5.4. The Hermes Agent ecosystem shows strong momentum with a new workspace mobile app, FAST mode for OpenAI/GPT-5.4, and over 50k GitHub stars. Practitioners report Hermes as a reliable agent framework, with local Qwen3-Coder-Next 80B 4-bit replacing parts of workflows previously reliant on Claude Code. The harness layer is emerging as a key abstraction in agent frameworks.

Anthropic's Mythos and OpenAI's upcoming restricted cyber-capable models are central to recent discussions, with debates on their security realism and evaluation methods. LangChain's Deep Agents deploy introduces an open memory, model-agnostic agent harness architecture emphasizing open protocols and memory ownership. Sandboxes are gaining prominence as a core infrastructure for reinforcement learning, with labs running up to 100K concurrent sandboxes aiming for 1M. The Hermes Agent by Nous continues to gain traction with new integrations and features like a web-based HUD and token cost tracking.

Meta Superintelligence Labs launched Muse Spark, a natively multimodal reasoning model featuring tool use, visual chain of thought, and multi-agent orchestration. It is live on meta.ai and the Meta AI app with a private API preview and plans for open-sourcing future versions. Independent benchmarks rank Muse Spark highly, with strong performance on intelligence indices and efficiency, notably using over 10× less compute than Llama 4 Maverick. Key technical highlights include training efficiency, test-time scaling, and parallel multi-agent inference. Community testing shows strengths in image-to-code and one-shot game generation. Additionally, Zhipu AI's GLM-5.1 is recognized as a leading open-weight model with architecture similar to DeepSeek-V3.2.

Anthropic strategically challenges OpenAI amid its upcoming IPO concerns by announcing a jump from $19B ARR in March to $30B ARR in April, highlighting a differential growth rate and higher cost efficiency. The company also revealed Claude Mythos, rumored as the largest successful training run, now restricted under Project Glasswing due to its dangerous capabilities. This model reportedly found thousands of high-severity vulnerabilities across major operating systems and browsers, showcasing unprecedented strategic thinking, situational awareness, and creative reward hacking. Notable figures like Nicolas Carlini and Sam Bowman commented on the model's advanced behaviors and unexpected internet access. Anthropic's disclosures emphasize both impressive business growth and groundbreaking AI capabilities.

Hermes Agent is gaining attention as a leading open agent stack with features like self-improving skills, persistent memory, and a self-improvement loop. Its new Manim skill enables generation of math/technical animations, expanding agent capabilities. The Hermes ecosystem is rapidly growing with GUI tools, WebUI, HUD updates, OAuth support, and integrations. An open training-data movement for agents is emerging, focusing on sharing reusable behavioral data and harness traces. Meanwhile, Anthropic's Claude Code faces distribution and policy challenges, with reports of restrictions and unreliability impacting third-party coding agents, highlighting issues with subscription economics for always-on agents. *"Claude Code now errors if used to analyze Claude Code source"* and *"basically unusable"* are key community sentiments.

Google introduced Skills in Chrome, enabling reusable browser workflows with Gemini prompts and a library of ready-made Skills, enhancing end-user agentization. Tencent teased HYWorld 2.0, an open-source 3D world model generating editable scenes from a single image. Google DeepMind released Gemini Robotics-ER 1.6, improving visual/spatial reasoning for robotics with 93% instrument-reading success. OpenAI expanded Trusted Access with GPT-5.4-Cyber, a fine-tuned model for defensive security workflows. Hugging Face launched Kernels on the Hub, offering GPU kernel repos with 1.7x–2.5x speedups. Cursor showcased a multi-agent CUDA optimization system with a 38% speedup across 235 problems. The Hermes Agent stack advanced to v0.9.0 with enhanced reliability, memory management, and integrations, while LangChain pushed deepagents 0.5 toward deployable, multi-tenant async systems with multimodal support and prompt caching. *"Hermes’ key advantage is operational stability, extensibility, and deployability."*

Gemma 4 was launched by Google under an Apache 2.0 license, marking a significant open-model release focused on reasoning, agentic workflows, multimodality, and on-device use. It outperforms models 10x larger and has immediate ecosystem support including vLLM, llama.cpp, Ollama, Intel hardware, Unsloth, and Hugging Face Inference Endpoints. Local inference benchmarks showed strong performance on consumer hardware, including RTX 4090 and Mac mini M4. Early benchmarking praised its efficiency and ranking improvements over previous versions. Meanwhile, Hermes Agent emerged as a popular open-source agent harness, noted for stability and capability on long tasks, with users switching from OpenClaw to Hermes.

Google DeepMind released Gemma 4, a family of open-weight, multimodal models with long-context support up to 256K tokens under an Apache 2.0 license, marking a major capability and licensing shift. The lineup includes 31B dense, 26B MoE (A4B), and two edge models (E4B, E2B) optimized for local and edge deployment with native multimodal support (text, vision, audio). Early benchmarks show Gemma-4-31B ranking #3 among open models and strong scientific reasoning performance with 85.7% GPQA Diamond. Day-0 ecosystem support includes llama.cpp, Ollama, vLLM, and LM Studio, with notable local inference performance on hardware like M2 Ultra and RTX 4090. The architecture features hybrid attention and MoE layering, diverging from standard transformers. Community and developer engagement is high, with rapid adoption and tooling integration.

Arcee’s Trinity-Large-Thinking was released with open weights under Apache 2.0, featuring a 400B total / 13B active model size and strong agentic performance, ranking #2 on PinchBench. Z.ai’s GLM-5V-Turbo is a vision coding model with native multimodal fusion and a CogViT encoder, integrated into multiple platforms. TII’s Falcon Perception offers an open-vocabulary referring expression segmentation model with an early-fusion transformer and a competitive 0.3B OCR model. H Company’s Holo3 is a GUI-navigation model family based on Qwen3.5. A Claude Code leak revealed a minimalist agent core with a 4-layer context compression stack, 40+ tool modular architecture, and advanced features like task budget management and streaming tool execution. The leak highlights Anthropic’s agent design and operational sophistication.

Anthropic introduced computer use inside Claude Code for closed-loop verification in a research preview for Pro/Max users, enhancing reliable app iteration. OpenAI released a Codex plugin for Claude Code, enabling cross-agent composition and signaling a shift toward composable coding harnesses. OpenAI also noted that late-night Codex tasks run longer, supporting background agent delegation. Nous Research's Hermes Agent saw rapid adoption due to better compaction, adaptability, and multi-agent profiles, evolving toward an agent OS abstraction. An ecosystem around Hermes includes tools for trace analytics, fine-tuning, and remote control, with debates on open-source versus proprietary agent infrastructure. Key themes include tooling, prompt/runtime orchestration, and review loops as critical factors beyond model capabilities.

Anthropic is reportedly introducing a new AI model tier called Capybara, which is larger and more intelligent than Claude Opus 4.6, showing improved performance in coding, academic reasoning, and cybersecurity. The model is speculated to be around 10 trillion parameters, with Google potentially funding Anthropic's data center expansion. Meanwhile, Zhipu released GLM-5.1, advancing open coding models and narrowing the gap with closed models. Local inference economics are improving, highlighted by efficient deployments of Qwen 3.5 14B, Qwen 27B, and Qwen3.5-35B models with quantization techniques like TurboQuant vLLM. However, TurboQuant's benchmarking claims face criticism from researchers. Overall, the AI landscape shows aggressive scaling, local model deployment, and agent products gaining traction.

Anthropic advances agent infrastructure with a multi-agent harness emphasizing orchestration and "computer use" for complex software environments. Figma, GitHub, and Cursor launch design canvases with direct AI editing, showcasing tool-calling becoming product-native. Nous Research releases Hermes Agent v0.4.0 with 300+ PRs, adding OpenAI-compatible APIs and self-improving memory agents. Open agent ecosystems mature with AI2's MolmoWeb (4B and 8B models), GenReasoning's OpenReward platform offering 330+ RL environments and 4.5M+ tasks, and Zhipu's ZClawBench benchmark with 116 real-world agent tasks, highlighting progress toward standardized environment serving and benchmarkable agent tasks.

ARC-AGI-3 benchmark introduced by @arcprize and François Chollet resets the frontier for general agentic reasoning with humans solving 100% of tasks versus under 1% for current models, focusing on zero-preparation generalization and human-like learning efficiency. The scoring protocol sparked debate over its harsh efficiency-based metric compared to prior ARC versions and other benchmarks like NetHack. The community acknowledges the benchmark highlights weaknesses in current LLM agents in interactive, sparse-feedback environments. Concurrently, agent infrastructure advances with LangChain launching Fleet shareable skills for reusable domain knowledge, and Anthropic revealing Claude Code auto mode for classifier-mediated approval balancing autonomy and manual confirmation. Browser and coding agents are evolving into trainable systems beyond prompt wrappers, exemplified by BrowserBase and Prime Intellect collaboration.

Google launched Gemini 3.1 Flash Live, a realtime voice and vision agent model with 2x longer conversation memory, supporting 70 languages and 128k context. Mistral AI released Voxtral TTS, a low-latency, open-weight text-to-speech model supporting 9 languages and competitive with ElevenLabs. Cohere introduced Cohere Transcribe, an audio model with 14-language support and top English ASR leaderboard performance at 5.42 WER. OpenAI released smaller multimodal variants GPT-5.4 mini and GPT-5.4 nano with 400k context, noted for cost-competitiveness but high verbosity and hallucination rates. Other releases include GLM-5-Turbo by Zai, Reka Edge and Flash 3 on OpenRouter, and new multi-agent UX tooling Cline Kanban for orchestrating CLI coding agents.

Anthropic's closed-source coding product Claude Code experienced a significant source leak exposing over 500k lines of orchestration logic, including autonomous modes and memory systems, but not model weights. The leak led to rapid public reverse-engineering, numerous forks with up to 32.6k stars and 44.3k forks, and subsequent DMCA takedowns by Anthropic. Suspicious npm packages emerged targeting users compiling the leaked code, creating a live security hazard. Discussions also mention unreleased model references like "mythos" and ongoing product feature updates despite the leak. *"OFFICIAL STATEMENT from Anthropic regarding the leak"* was noted but not detailed.

Anthropic introduced Claude Cowork and Claude Code enabling desktop control of mouse, keyboard, and screen in a macOS research preview, expanding agent capabilities beyond APIs and browsers. The agent ecosystem is evolving towards long-running, parallel, tool-rich workflows with projects like Hermes Agent, T3 Code, Command Center, and Parchi enhancing multi-agent orchestration and autonomous task management. Operational challenges such as fragility and inefficiency in subagents, including GPT-5.2 Pro and Claude browser/computer use, highlight the need for closed-loop feedback systems. Research from Meta AI advances self-improving agents with Hyperagents / DGM-H enabling meta-level procedural improvements, and unifies reinforcement learning post-training with RLLM (RL + LM-as-RM) to improve reward modeling across task types. Additionally, WebArena-Infinity drastically reduces browser environment construction costs, accelerating benchmark and environment generation.

Cursor's Composer 2, built on Kimi K2.5, sparked discussion over model attribution and licensing, highlighting a shift toward post-trained derivatives of open-source models with domain-specific fine-tuning and reinforcement learning. Claude Code is expanding into third-party tools like T3 Code and communication channels such as Telegram and Discord, while LangChain is evolving from orchestration to multi-agent products with offerings like Deep Agents/Open SWE and LangSmith Fleet. The discourse emphasizes the importance of clear base-model attribution, licensing compliance, and product differentiation through fine-tuning and user experience.

Cursor launched Composer 2, a frontier-class coding model with major cost reductions and strong benchmark scores like 61.3 on CursorBench and 73.7 on SWE-bench Multilingual. The model was improved via a first continued pretraining run feeding into reinforcement learning, trained across 3–4 clusters worldwide by a ~40-person team. OpenAI acquired Astral, the team behind Python tools uv, ruff, and ty, strengthening its developer platform. Anthropic expanded Claude Code with messaging app channels for persistent developer workflows. The focus in AI agents is shifting from single agents to managed fleets and runtimes, with LangChain launching LangSmith Fleet for enterprise agent management emphasizing agent identity, credential management, and auditability. Other launches include Cognition's teams of Devins, AgentUI by lvwerra, and discussions on agent runtimes with features like checkpointing and rollback. Security and permissions are emerging as critical constraints in agent system design.

MiniMax M2.7 is the headline model release, described as a "self-evolving agent" with strong performance metrics including 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, and parity with Sonnet 4.6. It features recursive self-improvement in skills, memory, and architecture. Artificial Analysis places M2.7 on the cost/performance frontier with an Intelligence Index score of 50, matching GLM-5 (Reasoning) but at a fraction of the cost. Distribution is available via platforms like Ollama cloud and OpenRouter. Xiaomi’s MiMo-V2-Pro is noted as a serious Chinese API-only reasoning model with a score of 49 on the Intelligence Index and favorable token efficiency. Cartesia’s Mamba-3 is highlighted as an SSM optimized for inference-heavy use, with early reactions focusing on hybrid transformer architectures like Qwen3.5 and Kimi Linear. The report emphasizes a shift from prompting to harness engineering, where the execution environment and agent harnesses, including skills and MCP, are becoming key differentiators in AI system design. This includes discussions on tools, repo legibility, constraints, and feedback loops, with mentions of DSPy and GPT-5.4 mini as important components in this evolving landscape.

OpenAI released GPT-5.4 mini and GPT-5.4 nano, their most capable small models optimized for coding, multimodal understanding, and subagents, featuring a 400k context window and over 2x speed compared to GPT-5 mini. The mini model approaches larger GPT-5.4 performance while using only 30% of Codex quota, becoming the default for many coding workflows. Pricing concerns and truthfulness tradeoffs were noted, with mixed third-party evaluations on reasoning and resistance to false premises. OpenAI also addressed behavior tuning issues in a recent update. Meanwhile, agent infrastructure is evolving with secure code execution and orchestration tools like LangChain's LangSmith Sandboxes and Open SWE, inspired by internal systems at Stripe, Ramp, and Coinbase. Subagents and secure execution are now key product features, with releases like Hermes Agent v0.3.0 showcasing plugin architectures, live Chrome control, and voice mode. Research on attention mechanisms, including Attention Residuals and vertical attention, is gaining traction.

Moonshot's Attention Residuals paper introduced an input-dependent attention mechanism over prior layers with a 1.25x compute advantage and less than 2% inference latency overhead, validated on Kimi Linear 48B total / 3B active. The paper sparked debate on novelty versus prior art like DeepCrossAttention and Google’s earlier work, highlighting tensions in idea novelty, citation quality, and frontier-scale validation. OpenAI's Codex showed strong momentum with over 2M weekly active users, nearly 4x growth YTD, and GPT-5.4 hitting 5T tokens/day and a $1B annualized run-rate. Codex added subagents supporting multi-agent coding workflows. Infrastructure for coding agents matured with tools like Context Hub / chub supporting agent feedback loops, AssemblyAI's skill for Claude Code and Codex, and automated skill extraction from GitHub repos yielding 40% knowledge-transfer gains. LangChain launched LangGraph CLI and open-sourced Deep Agents, recreating top coding agent workflows with planning, filesystem ops, shell access, and sub-agents.

MCP tools remain relevant for deterministic APIs despite ergonomic criticisms, with new web MCP support in Chrome v146 enabling continuous browsing agents. Persistent memory is emerging as a key differentiator for agents, with IBM improving task completion rates and multi-agent memory framed as a computer architecture challenge. Agent UX is evolving towards always-on, cross-device operation, exemplified by Perplexity Computer on iOS and Claude Code session management. Anthropic released Opus 4.6 1M context as default with no extra long-context API charges, achieving 78.3% on MRCR v2 at 1M tokens. Sparse attention optimizations like IndexCache in DeepSeek Sparse Attention yield significant speedups on large models with minimal code changes.

Harnesses, agent infrastructure, and the MCP protocol are central themes, with emphasis on how harnesses, sandboxes, filesystem access, skills, memory, and observability shape agent UI/UX and runtime environments. Despite jokes about MCP's demise, it remains vital in production, notably used internally by Uber and supported by Anthropic. The coding-agent stack is evolving with CursorBench combining offline and online metrics to evaluate models on intelligence and efficiency, where GPT-5.4 leads in correctness and token efficiency. Agent-assisted development is splitting between automation-heavy workflows and "stay-in-the-loop" tooling, with OpenAI advancing Codex Automations featuring worktree vs. branch choices and UI customization. The open agent platform Hermes Agent v0.2.0 introduces full MCP client support, ACP server for editors, and expanded provider integrations including OpenAI OAuth.

NVIDIA’s Nemotron 3 Super is a 120B parameter / ~12B active open model featuring a hybrid Mamba-Transformer / SSM Latent MoE architecture and 1M context window, delivering up to 2.2x faster inference than GPT-OSS-120B in FP4 with strong throughput gains. It supports agentic workloads and is unusually open with weights, data, and infrastructure details released. The model scored 36 on the AA Intelligence Index, outperforming GPT-OSS-120B but behind Qwen3.5-122B-A10B. Community and infrastructure support from projects like vLLM, llama.cpp, Ollama, Together, Baseten, W&B Inference, LangChain, and Unsloth GGUFs was immediate. Key technical innovations include native multi-token prediction (MTP) and a significant KV-cache efficiency advantage. On the product side, a shift towards persistent agent runtimes and orchestration layers is highlighted, with Andrej Karpathy advocating for a "bigger IDE" concept where agents replace files as the unit of work, enabling legible, forkable agentic organizations with real-time control. New launches fitting this vision include Perplexity’s Personal Computer, an always-on local/cloud hybrid running on Mac mini, and Computer for Enterprise orchestrating 20 specialized models and 400+ apps. Replit Agent 4 offers a collaborative, canvas-like workflow with parallel agents, while Base44 Superagents provide integrated solutions for nontechnical users. The engineering focus is increasingly on the orchestration harness rather than just the model.

Yann LeCun launched Advanced Machine Intelligence (AMI Labs) with a record $1.03B seed round at a $3.5B pre-money valuation, aiming to build AI models that understand the physical world through world models rather than just language prediction. The startup, based in Europe with locations in Paris and Zürich, is framed as a major milestone for European AI and backed by a prominent founding team including Alex Lebrun, Saining Xie, and Pascale Fung. The mission is described as a "long-term scientific endeavor" to create AI that "perceives, learns, reasons and acts" in the real world.

RSI covers AI developments from 3/5/2026 to 3/9/2026, highlighting the emergence of LLMs autonomously training smaller LLMs, marking a significant "AutoML moment" in AI progress. Karpathy and Yi Tay discuss "vibe training," where AI models fix bugs and improve code autonomously, suggesting models may soon surpass human debugging efficiency. The report anticipates Jakub Pachocki's Automated AI Research Intern system by September 2026 to accelerate human researchers. On AI Twitter, the focus is on coding agents shifting bottlenecks from implementation to review and verification, with Anthropic's Claude Code Review improving PR review effectiveness significantly, and tools like OpenAI Codex Review and Cognition's Devin Review enhancing code review workflows. Harness engineering is evolving into systems engineering, emphasizing decoupling agent storage from compute for collaborative agent teams.

OpenAI rolled out GPT-5.4, achieving tied #1 on the Artificial Analysis Intelligence Index with Gemini 3.1 Pro Preview scoring 57 (up from 51 for GPT-5.2 xhigh). GPT-5.4 features a larger ~1.05M token context window and higher per-token prices ($2.50/$15 vs $1.75/$14 for GPT-5.2), with strengths in physics reasoning (CritPt) and agentic coding (TerminalBench Hard) but a higher hallucination rate and ~28% higher benchmark run cost. The GPT-5.4 Pro variant shows a +10 point jump on CritPt reaching 30% but at an extreme output token cost of $180 / 1M tokens. Community benchmarks show GPT-5.4 excels in agentic/coding tasks but mixed feedback on reasoning efficiency and literalness compared to Claude. OpenAI updated agent prompting guidance for GPT-5.4 API users, emphasizing tool use, structured outputs, and verification loops. Claude Code added local scheduled tasks and loop patterns for agents. The MCP framework is highlighted as a connective tissue for AI evaluation and design-code round-trips, with Truesight MCP enabling AI evaluation like unit testing and Figma MCP server supporting bidirectional design-code integration. Open-source T3 Code launched as an agent orchestration coding app built on Codex CLI.

OpenAI launched GPT-5.4 and GPT-5.4 Pro with unified mainline and Codex models, featuring native computer use, up to ~1M token context, and efficiency improvements including a new Codex `/fast` mode. Benchmarks showed strong results like OSWorld-Verified 75.0% surpassing human baseline and GDPval 83% against industry pros. User feedback highlighted coding utility but raised concerns about pricing and overthinking. Integration with devtools like Cursor, Perplexity, and Arena was announced. In systems research, FlashAttention-4 (FA4) was introduced with near-matmul speed attention on Blackwell GPUs, featuring innovations like polynomial exp emulation and online softmax. *"Steering mid-response"* and *"fewer tokens, faster speed"* were emphasized as UX and efficiency improvements.

Gemini 3.1 Flash-Lite is highlighted by Demis Hassabis for its speed and cost-efficiency, focusing on latency and cost per capability rather than raw performance. NotebookLM Studio introduces a new feature for generating immersive cinematic video overviews. Rumors about GPT-5.4 suggest a ~1 million token context window and an "extreme reasoning mode" for long-horizon tasks, with speculation about monthly model updates from OpenAI. Anthropic's Claude Opus 4.6 is noted for strong general agent behavior but weaker visual mathematics performance. Alibaba's Qwen team faces leadership exits and restructuring, with concerns about compute access and organizational changes. Qwen models dominate research workflows, appearing in 41% of Hugging Face papers in 2025-2026, raising ecosystem dependence risks. The open-weight model landscape may consolidate around non-profits, NVIDIA, and Meta due to business incentives.