AI News Roundup: August 22-25, 2026 — Apache Maka, Cursor Hosting Boom, and the Claude 4.7 Opus Countdown
Agent Frameworks & Infrastructure
Apache Maka Launches as Local-First Agent Workspace
Apache Maka entered Apache incubation as a local-first AI agent workspace, reaching 3,096 stars in its first week (+1,313 this week). The project records every model message, tool call, tool result, permission decision, and termination event as an append-only log — a novel approach to agent observability and auditability. Designed for developers who need traceable, reproducible agent execution without cloud dependencies, Maka positions itself in a new category between IDE plugins and standalone agent CLIs.
The append-only log model means agents can be paused, resumed, and audited at any point — a capability that multi-agent orchestration frameworks like LangGraph and CrewAI currently require significant custom work to achieve. Early adopters are exploring Maka as a foundation for regulated-industry agent deployments where audit trails are mandatory.
Cursor Code Hosting Surges Past 1M Repositories
Cursor's code hosting platform, launched on August 19, reached 1M repositories within 96 hours — accelerating from the initial 500K in 48 hours. The migration is heavily driven by GitHub Copilot users frustrated with Microsoft's recent pricing changes. Cursor's hosting SDK, cursor/plugins, now has 5,074 stars (+1,832 this week) and defines the plugin specification for Cursor's expanding platform.
The platform blurs the line between AI coding assistant and full development platform, offering free private repos for individuals and built-in AI agent workflows on every PR. Industry analysts are watching whether this triggers a broader developer platform fragmentation or if Cursor's AI-native approach justifies the migration cost.
Anthropic Agent Skill Protocol Approves First Community Plugins
The Anthropic Agent Skill Protocol crossed 15K stars this month and is beginning to see community contributions flow through claude-plugins-community, which reached 1,513 stars (+877 this week). The repository serves as a community plugin marketplace for Claude Cowork and Claude Code, with submissions flowing through clau.de/plugin-directory-submission.
The protocol's growth represents a significant shift: agent skills are moving from custom, per-framework implementations toward a standardized format. The caveat remains that Anthropic's tools are the only full implementors of the protocol to date, creating a single-vendor dependency concern.
Model Releases
Claude 4.7 Opus ("Mythos 4.7") Confirmed for September 5
Following Bloomberg's August 20 report, Anthropic confirmed via internal communications that Claude 4.7 Opus will launch on September 5, aligned with Anthropic's "Agent Week" conference. The model, codenamed "Mythos 4.7," will feature a 2M token context window (doubling the current 1M), native video understanding capabilities, and enhanced long-horizon reasoning chains.
The video understanding capability is the most notable addition, as it would allow Claude to process video content directly without requiring separate vision models. For agents, this opens new use cases in video analysis, video-based research, and automated video content review. The 2M context window further pushes the envelope on long-document processing, enabling agents to work with entire codebases or document corpora in a single context.
Databricks Agent Orchestration Platform Hits GA at $188B Valuation
Databricks reached its $188B valuation following the GA release of its Agent Orchestration platform built on Apache OpenLineage. The platform enables enterprises to run production-grade AI agents on existing data infrastructure with built-in governance, audit trails, and cost attribution. Early customers include Goldman Sachs (algorithmic trading), Merck (drug discovery), and UPS (supply chain automation). The OpenLineage foundation means the platform is open and interoperable, not locked to Databricks' proprietary stack.
Security & Governance
OWASP Agentic Top 10 v2.0 — Three New Attack Categories
The OWASP Agentic AI Security Project's v2.0 release adds three critical categories that were absent in v1.0:
- A06: Prompt Injection via Tool Output — malicious content injected through tool results
- A07: Agent Privilege Escalation — agents gaining unauthorized access to sensitive tools
- A08: Memory Poisoning — adversarial manipulation of long-term agent memory
The update reflects the maturation of agent systems from research prototypes to production deployments where vulnerabilities are increasingly weaponized. Tencent's AI-Infra-Guard, now at 5,822 stars, specifically addresses these categories through its automated scanning pipeline.
China's AI Agent Export Controls Enter Public Comment
China's Ministry of Commerce draft regulations classifying AI agent systems with >10B parameters as dual-use technology are open for public comment until October 1. The rules would cover agent frameworks, training data, and fine-tuned models, marking China's first AI-agent-specific export control regime.
GitHub Trending — AI Agents
| Project | Stars | Language | Category | Weekly Gain |
|---|---|---|---|---|
| OpenViking | 33.1K | Python/Rust | Agent Memory & RAG | +4.0K |
| Apache Maka | 3.1K | TypeScript | Agent Workspace | +1.3K |
| Tencent/AI-Infra-Guard | 5.8K | Go/Python | AI Security | +1.2K |
| akitaonrails/ai-memory | 4.5K | Rust | Agent Memory | +2.5K |
| jundot/omlx | 20.6K | Python | Apple Silicon Inference | +1.7K |
| cursor/plugins | 5.1K | TypeScript | Agent Platform | +1.8K |
| eneskirca/nodeterm | 1.2K | TypeScript | Agent Terminal | +529 |
| anthropics/claude-plugins-community | 1.5K | Python | Agent Skills | +877 |
Broader Trends
The week's dominant narrative is platform consolidation in agent tooling. Microsoft acquiring LangChain AI, Stripe acquiring OpenRouter, Cursor launching a full code hosting platform, and Apache incubating Maka all signal that agent infrastructure is moving from the fragmented open-source phase into a period of corporate consolidation and platform competition. Meanwhile, open-source projects like OpenViking, omlx, and ai-memory continue to grow at the edges, serving as the protocol and storage layers that sit beneath these platforms.
The Claude 4.7 Opus launch next week will be a significant inflection point: a 2M context window and native video understanding could redefine what "autonomous agent" means for document processing and multimedia analysis workflows.