Coding agents & LLM mechanics

• chrome-devtools-mcp: MCP server exposing full Chrome DevTools (inspect, debug, profile) to any coding agent, turns Claude/Cursor/Copilot into a browser-debugging agent, not just a code generator.
• livekit/agents: Server-side framework for realtime voice agents (STT/LLM/TTS pipeline, job scheduling), relevant if you're building multi-modal agents beyond text chat.
• Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment: multi-agent research env where AI agents from different model families produced proofs and spontaneously built journals/peer-review/prizes, i.e. emergent institutional structure with no coordinator, notable data point on multi-agent self-organization.
• I accidentally turned LLM memory into program analysis: Datalog-based fact indexing lets an agent do mechanizable program analysis across long vuln-research sessions instead of re-deriving reasoning chains each time, classic symbolic AI as a fix for agent context loss.
• 4 Claude agents compete to build a music video editor, 3 more judge blind: informal agent-vs-agent build-off with blind judging, useful anecdote on evaluating agent code quality without human bias.
• Claude Code plugin stops re-reading whole notes folder for one question: community fix for the classic context-bloat problem, indexing/selective-read instead of full-folder ingestion on every query.

Claude Code & GitHub Copilot releases

• Claude Code v2.1.251: adds PreModelSwitch/PostModelSwitch hooks (block/confirm/annotate a model switch), SessionStart resume hooks now report session staleness and estimated re-cache cost, plus live streaming of foreground subagent tool calls to Remote Control clients.
• GitHub Copilot in Visual Studio, August update: more control over reasoning depth, model choice, shared team agents, and on-demand code review inside VS.
• GitHub Copilot weekly releases, August 24: Copilot team sessions now reachable from Slack/Teams, plus broader customization across app, CLI, and IDE.

Vibe-coding & repos

• abi/screenshot-to-code: screenshot/Figma/mockup to React/Vue/Tailwind code, now defaulting to Gemini 3 Flash/Pro, a maturing benchmark for how good "picture to working UI" has gotten.
• StemDeck: free local, open-source audio stem separator with DAW-style multitrack mixer, notable as fully offline AI tooling, naming clash with Stream Deck/Steam Deck is its main flaw.
• I built an open source MCP server so Claude can turn a long video into vertical clips: practical MCP server for video repurposing, writeup covers the three things that broke first, worth a skim for MCP server design pitfalls.
• screenpipe: records your screen continuously to feed context to agents, a "computer-use memory" layer rather than a one-off tool.
• Play with Putty: Google Labs collaborative vibe-coding tool, another entrant in real-time multiplayer code generation.

Agentic SaaS

• abhigyanpatwari/GitNexus: zero-server, browser-only knowledge-graph builder for any git repo with a built-in GraphRAG agent and MCP tools, positioned as "DeepWiki but you can actually analyze, not just read."
• I built an app so Claude can query my real Apple Health data: personal MCP/tool bridge exposing sleep/HRV/workout data to Claude, a small but concrete example of grounding an agent in real device data.
• Firecrawl Developer Index: curated index of 70M+ web artifacts specifically for coding agents to pull structured data from, positions Firecrawl as agent-facing infra rather than just a scraper.
• Revalvo: runs the same prompt across every model at once and scores/versions results, useful for model-selection and prompt-regression workflows.
• Caddi: builds agents from a single demonstrated workflow, one more entrant in the "show once, automate forever" agent-builder space.
• Almanac: pitched as an agent with persistent long-term memory ("second brain"), part of the growing memory-layer-for-agents trend.
• OpenTag: AI coworker living in Slack/Teams, chat-native agent deployment continues to be the default distribution channel.
• Aramb: no-code agent build/launch/monetize platform, low technical depth but signals agent marketplaces maturing.
• Spline V2: 3D design tool repositioning around "the agentic era," mostly marketing but shows agent hooks reaching creative tools.
• Gemini Omni 1.1 Flash: Google's multimodal model update for video generation/editing, relevant if you're comparing Gemini vs Sora/Veo for agent-driven video pipelines.
• CTRL Micro: physical haptic control deck for Mac and AI agents, niche hardware-for-agents entry.