The agent stack is splitting cleanly into models you rent and runtimes you own. Meta shipped Muse Glimmer, a 30-billion-parameter open-weight model built for local agentic workflows on consumer hardware. SpaceX finalized its $60 billion all-stock acquisition of Anysphere, folding Cursor into the rocket company's AI orbit. Meanwhile, Microsoft, DeepSeek, and a cross-vendor coalition each open-sourced the agent harness layer - Agent Lightning v1.0, DeepSeek Harness v0.1, and Agent Plugins 1.0 - turning the runtime that wraps the model into portable, hackable infrastructure.
AI & ML News
Meta Releases Muse Glimmer, a 30B Open-Weight Model for Local Agentic Workflows
Meta Superintelligence Labs published Muse Glimmer on August 10, 2026, a dense 30-billion-parameter multimodal model under Apache 2.0 on Hugging Face. The model was distilled from Muse Spark, Meta's closed frontier model, and targets agentic tasks - tool use, planning, and multi-step execution - that run on a single consumer GPU or Apple Silicon Mac. Meta's benchmarks show Glimmer beating Qwen3.6-27B on most agentic evaluations and outperforming Gemma 4-31B across the board. CEO Mark Zuckerberg framed the release as a push to lower U.S. barriers on open-weight AI so American developers can compete with Chinese rivals, and promised larger open releases soon. The model is available for download without API keys or usage restrictions.
SpaceX Completes $60 Billion All-Stock Acquisition of Cursor Maker Anysphere
SpaceX closed its acquisition of Anysphere, the company behind the Cursor AI code editor, on August 14, 2026, in an all-stock transaction valued at $60 billion - the largest venture-backed startup acquisition on record. Anysphere shareholders received SpaceX Class A shares priced on a seven-day volume-weighted average. The deal was announced June 16, days after SpaceX's Nasdaq debut at a $2 trillion-plus valuation. Cursor's four MIT-educated cofounders - Michael Truell, Aman Sanger, Sualeh Asif, and Arvid Lunnemark - each stand to reach roughly $2.7 billion in net worth upon closing. Cursor reports approximately 4 million active developer users and $2.6 billion in annualized revenue. SpaceX merged with xAI earlier in 2026, and the Cursor acquisition is expected to feed the combined entity's Colossus training supercomputer and Grok coding agent roadmap.
Google Releases Gemini 3.7 Flash, Its Most Capable Workhorse Model for Coding and Agents
Google made Gemini 3.7 Flash generally available on August 13, 2026, billing it as a substantial improvement across software engineering, web development, and agentic workflows. The model is available at an introductory price through December 31, 2026, via the Gemini API and in the Gemini app. It follows the July 21 release of Gemini 3.6 Flash and 3.5 Flash-Lite, and adds to a 3.x generation that now spans Flash, Flash-Lite, Pro preview, and Deep Think reasoning modes. Google also rolled out an updated Gemini 2.5 Pro I/O edition with enhanced coding capabilities on August 19, and a dedicated student hub with study notebooks on mobile. The 3.7 Flash release is the latest signal that Google is optimizing the Flash tier for the agent workloads that now drive the majority of API traffic.
OpenAI Retires o3 from ChatGPT on August 26 After 90-Day Sunset
OpenAI will remove o3 from the ChatGPT model picker on August 26, 2026, concluding a 90-day sunset period announced on May 28. The reasoning model, launched April 16, 2025 alongside o4-mini as the company's "smartest models released to date," will remain available via API snapshots until December 11, 2026, when the gpt-5.6-sol replacement takes over. The retirement applies only to ChatGPT web and mobile; API access is unaffected until the December snapshot shutdown. OpenAI cited retiring "older models with limited usage" to support newer GPT-5-series models. The move completes the phase-out of the GPT-4 era in ChatGPT, which began with GPT-4o and GPT-4.1 retirements in February 2026 and GPT-4.5 in June.
Microsoft Agent Plugins 1.0 Standard Goes GA Across VS Code, Copilot, Cursor, and ChatGPT
A coalition of Amazon, Microsoft, OpenAI, Vercel, and Cursor adopted Agent Plugins 1.0 on August 6, 2026, an open, vendor-neutral standard for packaging agent skills and MCP servers into portable plugins. Google joined as a core maintainer the same day. The standard is now generally available in VS Code, GitHub Copilot CLI, the Copilot SDK, the Copilot app, and ChatGPT. Agent Plugins 1.0 deliberately excludes installation mechanisms, distribution protocols, provenance verification, permission models, sandboxing, and marketplaces - leaving each platform to build its own distribution channel. The Technical Steering Committee includes Clare Liguori (AWS), Roshan Sadanani (Cursor), Harald Kirschner (Microsoft), Gav Verma (OpenAI), and Jonathan Hefner (Vercel) as lead maintainer.
Big GitHub Releases
Microsoft Agent Lightning v1.0: Harnessed Agentic RL in ~3,500 Lines
Microsoft Research released Agent Lightning v1.0 on August 17, 2026 (arXiv:2608.17528), a lightweight MIT-licensed framework for reinforcement learning on LLM agents that trains the policy model through the same harness used in production. The disaggregated architecture bridges arbitrary agent harnesses - Claude Code, OpenCode, Cursor, OpenHands, mini-SWE-agent - to RL backends like verl, handling retokenization, sample merging, rollout-level advantage computation, and loss normalization. On SWE-bench Verified, RL with Agent Lightning improved Qwen3.5-9B from 41.8% to 56.4%, a 14.6-point absolute gain using only ~6,000 training examples and modest compute. The framework includes open data pipelines, reward-hacking safeguards, and reproducible training scripts.
DeepSeek Harness v0.1: "Everything Is a Plugin" Agent Runtime
DeepSeek open-sourced DeepSeek Harness v0.1 on August 13, 2026, an MIT-licensed agent harness built on the Cordis meta-framework where models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and the UI are all swappable plugins. The developer preview scores 87.9 on Terminal-Bench 2.1 with the newly released DeepSeek-V4-Pro-0813 model (up from 72.1 on the April preview). DeepSeek warns of compatibility-breaking changes and positions v0.1 as an experimentation target, not a production pin. The release coincides with API price increases effective August 16: V4-Pro input rises from $0.435 to $0.66 per million tokens off-peak, output from $0.87 to $1.98, with peak rates double during Chinese working hours.
GitHub Trending: Agent Skills Frameworks Dominate the Leaderboard
The GitHub trending page for August 22-23 is led by agent-skills ecosystems. mattpocock/skills sits at 232,469 stars with 1,894 stars gained today, a shell-based collection of engineering skills for Claude Code, Codex, and Cursor. obra/superpowers reaches 276,287 stars (557 today), an agentic skills framework and methodology. mukul975/Anthropic-Cybersecurity-Skills holds 30,697 stars with 766 today, offering 817 production-grade cybersecurity skills across 29 domains mapped to MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3. These repositories represent the standardization layer turning raw model capability into repeatable, auditable agent behavior.
GitHub Copilot Adds Grok 4.6, Gemini 3.7 Flash, and Kimi K3 in August Rollout
GitHub's August 10 weekly Copilot release added Grok 4.6, Gemini 3.7 Flash, and the cost-effective Kimi K3 model (billed at $3/$15 per million input/output tokens) to the model roster. Agent Plugins 1.0 reached general availability across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app. Per-model token breakdown appeared in usage reports on August 11, letting enterprises track spend by model. MAI-Code-1.1-Flash gained native image understanding, and the upcoming September 1 deprecation of MAI-Code-1-Flash was confirmed. GitHub also launched shared agentic work in Microsoft Teams and a new Slack experience for Copilot.
Under-the-Radar Finds
akitaonrails/ai-memory, 4,118 Stars, Long-Term Memory for Agent Coding CLIs
akitaonrails/ai-memory solves the handoff problem between different agent vendors by providing persistent, portable memory for coding agents. The Rust-based system lets agents store and retrieve context across sessions and across tools - Claude Code, Codex, OpenCode, and others - so switching agents doesn't mean losing project knowledge. It gained 606 stars on August 22 alone. As multi-agent workflows become standard, a vendor-neutral memory layer is the missing piece that lets developers treat agents as interchangeable workers rather than locked-in platforms.
nautechsystems/nautilus_trader, 27,338 Stars, Production-Grade Rust Trading Engine
nautechsystems/nautilus_trader is a deterministic, event-driven trading engine written in Rust with 27,338 stars and 80 stars gained today. It supports backtesting, paper trading, and live execution across crypto, equities, and futures with a unified API. The architecture enforces strict event ordering, nanosecond timestamps, and zero-allocation hot paths - features that matter when agents start executing financial workflows autonomously. As AI agents move from code generation to capital allocation, a battle-tested, open-source execution engine becomes critical infrastructure.
jundot/omlx, 20,367 Stars, LLM Inference Server for Apple Silicon with Menu Bar Management
jundot/omlx runs LLM inference with continuous batching and SSD caching on Apple Silicon, managed from the macOS menu bar. At 20,367 stars and 472 gained today, it fills a gap for developers who want local model serving without Docker or Kubernetes complexity. The server supports popular open-weight models and exposes an OpenAI-compatible API, making it a drop-in replacement for cloud endpoints during development. For the growing cohort of developers running models locally on Macs, omlx is becoming the default control plane.
santifer/career-ops, 67,814 Stars, Open-Source AI Job Search With Structured Evaluation
santifer/career-ops scans job portals, evaluates listings with a structured A-F rubric into a 1.0-5.0 score, tailors CVs, and tracks applications - all running locally inside your AI coding CLI (Claude Code, Codex, OpenCode, Antigravity). At 67,814 stars and 198 gained today, it demonstrates how agentic workflows can automate high-stakes personal tasks without sending data to a SaaS platform. The project uses a local-first architecture where the LLM calls happen inside the user's chosen CLI, keeping PII on the machine.
mukul975/Anthropic-Cybersecurity-Skills, 30,697 Stars, 817 Skills Mapped to Six Security Frameworks
Mukul975's Anthropic-Cybersecurity-Skills is the largest open-source cybersecurity skills library for AI agents, with 817 production-grade skills spanning 29 security domains mapped to MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3. It runs across 26-plus agent platforms including Claude Code, Codex CLI, Cursor, and Gemini CLI, and ships under Apache 2.0 with authorized-use guardrails covering red-team command and control, phishing simulation, and exploitation. For security teams wiring agents into their workflows, it is a turnkey curriculum rather than a blank slate.
Why it matters. August 23, 2026 shows the AI stack stratifying: models (Muse Glimmer, Gemini 3.7 Flash, GPT-5.6) are becoming commoditized utilities you rent, while the agent runtime layer (Agent Lightning, DeepSeek Harness, Agent Plugins 1.0, herdr, codebase-memory-mcp) is where open-source competition is fiercest and lock-in is weakest. SpaceX buying Cursor for $60B signals that the valuations now attach to the distribution surface - the editor, the harness, the plugin marketplace - not the model weights. The winners will own the layer developers touch daily.