Coding agents & LLM mechanics

• OpenAI whitepaper on treating Codex as persistent workspace — context preservation across long-running multi-session work, relevant to how you'd structure autonomous goal runs.
• GPT-5.6 (Sol/Terra/Luna) launches publicly Thursday, broad preview rollout starting now.
• Early takes on GPT-5.6: strong at computer use, fixed front-end design gaps and runs unattended for a day without /goal, praised as best model tried, front-end finally fixed — all third-party hype, no benchmarks yet.
• Anthropic's "J-Space" paper describes an internal workspace where Fable holds concepts before responding; someone built a skill to expose it (skill download) — caveat noted it's self-report, not raw activations.
• Matt Pocock's proxy-and-trim method cuts Claude Code system prompt from 21.6K→13.6K tokens (~37% saved per turn) by intercepting what actually gets sent and stripping cruft via settings.json — directly reusable for your own CC setup (thread start, replicated result, ask for harness-agnostic proxy tool).
• Claude Code ships Artifacts: interactive shareable pages (PR walkthroughs, dashboards) from a session, beta on Team/Enterprise.
• Fable 5 access on paid plans extended through July 12.
• Matt Pocock teasing Skills v1.1: /wayfinder, /to-spec, /to-tickets, Martin-Fowlerized /code-review, /triage on external PRs — shipping this week.
• GLM 5.2 called incredible but not truly self-hostable / not comparable to Fable — pushback on positioning claims.
• Theo's stance: review much less of your code than 5 years ago — trust-the-agent argument, worth weighing against your own review habits.

Vibe-coding / agent-driven builds

• levelsio had Claude Code build+run an iOS Swift app via SSH into a rented MacinCloud Mac Mini from his Hetzner VPS — VPS-only coding workflow, cloud Mac as build target (setup thread).
• Same pattern for automating UberEats orders: Claude Code + Playwright, login once, then natural-language commands.
• Fable used to write an original NES game in 6502 assembly from scratch, runs within real cartridge constraints — stress-tests low-level/constrained-target code generation.

Agentic SaaS / infra

• MagicPath 2.0: multiplayer canvas where humans + Codex/Claude Code co-design/build prototypes live.
• agents-cli: ships AI agents onto Google Cloud, driven from your coding agent.
• NanoKVM-Go: gives an AI agent physical KVM control over any screen — hardware-level computer-use.
• Uber: 99% of engineers use AI tools, 70%+ PRs agent-attributed, 2,500+ internal agent skills — large-scale agentic adoption data point.