Coding agents

• Minimal-pair study claims code cleanliness doesn't affect agent task completion — but community pushback notes "clean" condition used AI-cleaned repos, not human-organized ones, undercutting conclusion (arxiv).
• pxpipe cuts Claude Code input tokens by rendering bulky context (system prompt, tool docs, history) as images — dense text packs ~3.1 chars/image-token vs ~1 char/text-token, same vision channel reads it back.
• codebase-memory-mcp indexes repos into persistent knowledge graph via tree-sitter across 158 languages — Linux kernel (28M LOC) in 3min, sub-ms structural queries, 99% fewer tokens, single static binary.
• 130+ page survey on always-on agents — systems whose future behavior depends on durable state (task ledgers, permissions, credentials, provenance, triggers) beyond simple memory.
• Claude Code ships Artifacts: interactive pages from a session (PR walkthroughs, dashboards) shareable via private link, beta on Team/Enterprise — early user take says it's become daily-driver for code visualization and diagrams.
• Codex getting "Ultra" tier per OpenAI eng tweet — no detail yet.
• Bindu Reddy argues no single model wins chat+code — future is prompt-routed mixture-of-agents by intent.
• Meituan LongCat-2.0: 1.6T params (~48B active), 1M context, MOPD arch (agent/reasoning/interaction experts), claims SWE-bench Pro 59.5 > Gemini 3.1 Pro/GPT-5.5/Opus 4.6 — to be open-sourced.
• Interconnects: frontier post-training recipe review w/ Finbarr Timbers on getting an Olmo-style recipe to frontier quality.
• Interconnects: Claude Fable 5 safety fables — Anthropic quietly classified/steered "AI safety" research queries, later added a classifier akin to other safety domains after pushback.
• System prompt leaks repo — extracted prompts from Claude/ChatGPT/Gemini/Grok/Cursor/Copilot etc, now diffing Opus 4.8 → Fable 5 changes.
• Leaked GPTs repo, 32k stars, publishes hidden system prompts for 200+ custom GPTs, still up.

Agentic SaaS

• MagicPath 2.0 — multiplayer canvas where humans + agents (Codex, Claude Code) co-design/build, pulls codebase/data context, live agent view.
• SnapState — persistent state layer for AI agent workflows.
• TryCase — disposable test environments for AI coding agents.
• CircleChat — Slack + task board + "boss" for AI agents to report into.

Vibe-coding & repos

• OpenMontage — open-source agentic video production: 12 pipelines, 52 tools, 500+ skills, turns coding assistant into full video studio.
• ai-job-search — Claude Code framework: fit-eval, CV tailoring, cover letters, interview prep via drafter-reviewer pipeline.
• T3MP3ST — open-source multi-agent red-teaming meta-harness, automates recon/vuln-discovery/analysis for authorized engagements (tweet).
• macOS YouTube-TV menu-bar app one-shot via Codex /goal — "golden age of ideas."
• Play Store asset creation automated via Claude Code skill, open-sourced.
• AI Coding Cohort v2 — 2500+ students, Claude Code + AFK agents + Sandcastle, starts June 1.
• Anecdote: fully vibe-coded profitable SaaS by non-coder, framed as question/provocation.