Coding agents

• Fable 5 vs. GPT-5.6 Sol on NP-hard problem: head-to-head test, /goal mode helps in long sessions, not this short benchmark, single-run noise so weak signal.
• Claude v2.1.215 release: /verify and /code-review no longer auto-run, must invoke explicit now, changes default workflow behavior.
• Fable 5 permanent: Fable 5 rolls into all Max/Team Premium plans July 20 at 50% limits, Pro/Team Standard stay credit-based.
• Skill engineering vs one-shot design: case for treating skills as engineered artifacts, human judgment still needed to steer agents even as capability rises.
• Autoresearch self-improving loops: "outer loop" pattern where agents maintain/improve their own recipes, humans stay central to the factory.
• AIEWF loops debate wrap: conference-closing debate on agent loop architectures, state-of-AI-engineering snapshot.
• Vercel's Andrew Qu on agents as new software: Vercel's eve framework rationale, skills/sandboxes/agent-readable sites framed as core infra now.
• GLM-5.2 step change for open agents: open-weight model crosses capability threshold for agentic use, paired paper on open RL recipes for terminal agents.
• Harness Engineering explainer: argues system around model (harness) now drives more gains than prompt/model swaps alone.
• Distillation panic take: pushback on "distillation attack" framing, conflates legit distillation with API-scraping jailbreaks.
• Spare Mac as always-on Claude Code box: step-by-step for 24/7 computer-use setup reachable via phone/SSH, HN thread skeptical on killer app.
• Burned out merging AI code, built Claude skill: reddit skill to ease AI-code merge fatigue.
• AI Coding Cohort v2: Matt Pocock's course update, agent-agnostic, adds AFK agents via Sandcastle, starts June 1.

Vibe-coding & repos

• stablyai/orca: ADE running Codex/Claude Code/OpenCode/Pi in parallel worktrees, one dashboard, usage tracking across providers.
• mattpocock/skills: personal .agents skill library, model-agnostic, explicit anti-"process ownership" stance vs GSD/BMAD/Spec-Kit.
• ticketmill Claude Code plugin: batch-processes GitHub issues end-to-end with human merge gate.
• LangChain's OSS agent factory: full internal SWE-agent stack open-sourced, blog breaks down each component and why OSS matters.
• Synapse local codebase indexer + MCP: local indexer exposed as MCP server for Claude Code.
• KnockOutEZ/wigolo: local-first web search/fetch/crawl MCP tool, no API keys, no cloud cost, works across major agent clients.
• ogulcancelik/herdr: terminal agent multiplexer, real per-agent views, detach/reattach over SSH, socket API for agents to spawn/read own panes.
• 4-model SaaS build stunt: 200+ tasks, zero supervision, 3 days agentic coding across 4 models, framed around "shipping is the hard part" thesis.

Agentic SaaS

• MagicPath 2.0: multiplayer canvas for humans + Codex/Claude Code agents to co-build prototypes live, pulls data from any source.
• Clark: AI coworker with its own persistent cloud computer.
• ZooData: positions itself as "data layer for AI agents."
• SnapState: persistent state layer for AI agent workflows.