Google releases Gemini 3.8 Live and Extended Thinking for developers to build real-time voice agents that execute tasks mid-conversation, adding asynchronous function calling, live visual context and support for more than 97 languages.
Salesforce launches Koa, a business reasoning model built on Nvidia Nemotron 3 Super for long-running AI agents, alongside AIforce, Claudeforce’s 37 pre-built sales skills, and Headless 360 API access at Dreamforce 2026.
Gitlawb launches OpenClaude, an open-source terminal coding agent that works with cloud APIs and local AI backends, including Gemini, Ollama, Bedrock and OpenAI-compatible providers, while its --bg mode runs prompts in detached sessions with logs stored in ~/.openclaude/bg-sessions/.
Chat On Steroids now gives ChatGPT local MCP access to edit approved files, run tests, retain terminal sessions and control desktop tools through Chrome, while supporting multi-agent workflows on Windows, macOS and Linux.
Google releases ARTEMIS, an open-source platform that converts natural-language requests into Android automation, linking real devices to Antigravity, Codex, Claude Code and Windsurf via MCP, with a Pro mode designed for workflows of more than 100 steps.
shadcn-ui launches @shadcn/lint, an agent-first Tailwind linter for ESLint and Oxlint that guides coding agents to fix component, variant and theme violations in one correction round, with support for Tailwind v4 and Node.js 20.19+.
DreambigOu releases ELI5, a Claude Code skill that tailors explanations for children, managers, engineers and parents, with its evaluation reporting an 83.3% pass rate versus 41.6% without the skill.
LLMTraceFX ships a deterministic KV-cache auditor in merged PR #74 that independently verifies cache attestations and flags evictions, confirming token-identical cached outputs across 10 synthetic tests while showing an exact duplicate reuses 8 of 9 tokens.
LibreChat v0.8.8-rc3 launches a beta Agent Management API and GPT-6 Astra support, while introducing experimental agent workspaces that can edit files and run Bash commands under bounded timeouts.
Google DeepMind launches the mujoco-menagerie Python package for on-demand curated robot simulation models, letting users load Unitree Go2 with mm.load('unitree_go2') and pin models to the 2026.9.0 release.