Coding agents

• Claude Code v2.1.199 ships stacked slash-skill invocations (up to 5 chained) and fixes SSL cert failures that used to burn retries silently — release notes.
• claude --bg --name "Session Name" "prompt" spins up a named background agent for clean handoff between sessions — tip.
• Artifacts in Claude Code (write, publish live to claude.ai, update in real time) now rolled out to Pro/Max, not just Max-only preview — announcement.
• T3 Code reportedly has a working branch letting Codex subagents spawn from Claude sessions and vice versa — cross-vendor agent delegation getting normalized — tweet.
• GitHub Copilot agent session streaming hits public preview for Enterprise Cloud — full session data across Copilot clients, useful for audit/observability — changelog.
• openai/codex-plugin-cc brings Codex reviews (/codex:review, /codex:adversarial-review) and task delegation directly into Claude Code — repo.

LLM mechanics & token tricks

• One user measured ~10.5M tokens/week wasted by Claude Code re-reading files and command output — concrete data on why caching/context hygiene matters — reddit.
• caveman skill cuts ~65-75% of output tokens by forcing terse "caveman speak" while preserving technical accuracy — works across Claude Code, Codex, Cursor, Windsurf, 30+ tools — repo.
• "Short leash" method argues year-long research shows tight, prompt/skill-constrained AI agent use beats loose autonomy for security-critical code — HN pushback notes handwritten code still wins on efficiency — post.
• Paul Bakaus makes the case for "skill engineering" over one-shot AI design — durable skills need human judgment in the loop, not full autopilot — latent space.
• agentskills/agentskills proposes an open, lightweight spec (SKILL.md + metadata) for portable agent capabilities across tools — repo.
• AI Engineer World's Fair closed on a "loops" debate — core tension of the conference: how much agentic autonomy vs. structured control is right for production AI eng — dispatch.
• Simon Willison used DSPy to systematically evaluate/improve Datasette Agent's SQL system prompts — practical example of prompt optimization via structured eval rather than vibes — writeup.

MCP & browser tooling

• chrome-devtools-mcp gives coding agents (Claude, Cursor, Copilot) live control of Chrome DevTools via MCP — debugging, perf analysis, automation without screenshots — repo.
• WebKit ships a Safari MCP server in Safari Technology Preview 247, giving agents direct access to Safari's dev/debug workflow — blog.
• Manufact (YC S25) launches MCP Cloud — hosting, cross-client testing (ChatGPT/Claude), traffic inspection/replay for MCP servers — HN skeptics question if it's more than CLI + config — YC launch.
• PieterPost MCP connects agents to physical postal mail — niche but signals MCP creeping into real-world action surfaces — producthunt.

Vibe-coding & repos

• affaan-m/ECC is an agent-harness performance system (skills, instincts, memory, security, research-first dev) for Claude Code, Codex, Opencode, Cursor — ships official-source-only warnings against third-party re-uploads — repo.
• craft-ai-agents/craft-agents-oss is craft.do's open-sourced tool for document-centric (not code-centric) multi-agent sessions — repo.
• santifer/career-ops is a Claude-Code-built job-search system: 14 skill modes, Go dashboard, PDF gen, batch processing, fully open-sourced — repo.
• Simon Willison's llm-coding-agent is an early experiment building a minimal coding agent on top of his llm library, now evolved into an agent framework — release.
• claude-real-video claims to let Claude "watch" video frame-by-frame rather than transcript-only — community pushback: still fails on motion/animation inference, and frames do get sent to Anthropic despite privacy claims — repo.

Agentic SaaS

• Scritty pitches shared, searchable memory across every AI coding agent you use — cross-tool context persistence as a product category — producthunt.
• Retrace lets you debug AI agents by replaying and forking their runs — observability tooling for agent failure analysis — producthunt.
• Vercel's Chief of Software Andrew Qu explains eve, their internal agent framework, and argues skills/sandboxes/agent-readable websites are the new infra primitives — interview.