Foothills of the Singularity

What we talked about

• Skills 101 — what a Claude Code skill is, what goes in SKILL.md, and how the harness decides when to load one. Each skill is a folder with SKILL.md: YAML frontmatter at the top (name + description), Markdown body below. Only the description loads into context by default — the harness reads the description on every turn and, when something in the conversation matches, pulls the body in on demand. Cheap to add, free when unused — your prompt only pays for the skills it actually invokes. Walked through prepare, news, and schedule from this repo as concrete examples.
• PI agent — sustainably as a separate service (earendil-works/pi) — Mario Zechner’s (badlogicgames) open agent harness as an answer to “how do I reproduce a Claude-Code-style loop sustainably as a service I control?” Monorepo of pi-coding-agent (interactive CLI), pi-agent-core (tool-calling runtime with state management), and pi-ai (unified multi-provider LLM API across OpenAI / Anthropic / Google), plus a TUI library and a separate pi-chat Slack bot. Self-extensible — the agent can edit and explain itself — and Mario publishes his own pi-mono work sessions to Hugging Face (badlogicgames/pi-mono) as real-world training data for OSS coding agents.
• Playwright play — using the Playwright MCP server to let Claude drive a real browser end-to-end: load the page, click around, fill forms, screenshot, assert against the DOM. Closes the loop on the CLAUDE.md rule that UI / frontend changes get exercised in a browser before they’re called done — instead of asking the human to verify, Claude takes the screenshot itself and reads it back. Pairs cleanly with claude agents for headless verification runs against a dev server.
• OpenHands — the open-source autonomous software-engineer agent (formerly OpenDevin, out of All Hands AI). Runs each task inside a sandboxed Docker runtime so the agent can edit files, run shells, and drive a browser without touching your host. Model-agnostic — point it at Claude, GPT, a local model — and pluggable via microagents / MCP. The natural reference point when comparing “Claude Code + skills” against an end-to-end open agent harness with its own UI.
• Custom statusline (docs) — bottom-of-CLI status bar driven by a shell script you point at via statusLine.command in settings.json. Claude Code pipes a fat JSON blob to stdin every assistant turn (model, cwd, git repo + worktree, cost, context-window %, 5h / 7d rate-limit %, vim mode, open PR + review state, agent name, …) and prints whatever your script writes to stdout. Multi-line output, ANSI colours, OSC 8 clickable links, optional refreshInterval for time-based segments. Runs locally — costs no tokens. Or just run /statusline and describe what you want.

Announcements

• OKTech — Kitchen Robots & State of AI — Sat 30 May, 17:00 ~ 19:30, Cybozu Osaka (35F Hankyu Bldg). Robotics in the home + AI in web dev. — oktech.jp

Headlines

1. Google I/O 2026 — Demis’s “foothills of the singularity” AGI framing (agents + multimodal + robotics + scientific AI, 2030 ± 1); Antigravity CLI replacing Gemini CLI with a June 18 deadline; Gemini Omni’s physics-grounded video as a step toward general-purpose robotics; Gemini 3.5 Flash GA (4× faster, planned as the default behind everything Google); native Android in AI Studio; Google AI Ultra at $100/mo. What actually changes for engineers in this room? — Semafor — foothills quote — Antigravity CLI transition — Gemini Omni — I/O developer recap — Simon’s I/O analysis
2. Managed Agents — Anthropic and Google — both shipped “single API call → reasoning agent in a sandboxed Linux env with persistent state” the same week. Anthropic added MCP tunnels (route to private MCP servers without exposing them) and self-hosted sandboxes (public beta) on May 19. Compare surfaces, isolation, billing, what developers actually have to think about. — Google Managed Agents — Anthropic — MCP tunnels + self-hosted sandboxes

Robert Demo

• TBD.

Discussion Topics

1. How are you using coding agents? — opener. Which tools, what shipped, what got dropped, where AI hit a wall this week.
2. Claude alternatives — Cline SDK + OpenCode SQLite — Cline extracted its agent harness into @cline/sdk (open-source TS, one harness across VS Code + JetBrains + CLI). OpenCode v1.2 swapped flat-file sessions for SQLite — and silently orphaned existing JSON sessions on incremental upgrades. Is the open-source agent stack reaching parity? — cline/cline — OpenCode v1.2.0
3. /code-review (Claude Code) — /simplify → /code-review with effort levels and --comment for inline PR review. Where does this fit alongside gh pr review and our existing review loops? — v2.1.147 release

Show and Tell

Member picks

• PromptZero — openbashok/promptzero — local proxy that anonymises sensitive data in prompts before they reach Claude. IPs, hostnames, emails, credentials, national IDs (AR / CL / ES / UY / CO / MX), phone numbers, credit cards, API tokens, person/org names — across English and Spanish via NLP plus regex. Real values get swapped for IANA documentation ranges (RFC 5737 / 3849 / 2606); session-scoped mapping tables restore them on the response. Pair with Burp Suite or mitmproxy to audit that only sanitised data leaves your box. — github.com
• Gemini Omni — Google’s new multimodal “world model” — text + image + audio + video in, editable physics-grounded video out. Demis frames physics-grounded video as a step toward general-purpose robotics; the practical bit today is video editing via natural-language conversation, SynthID-watermarked. Omni Flash rolls out to AI Plus / Pro / Ultra via the Gemini app and Google Flow, plus free on YouTube Shorts and YouTube Create. — blog.google
• Antigravity CLI — Google’s replacement for Gemini CLI, shipped at I/O 2026 alongside the desktop app rebuild. Terminal-first agent spawning, same harness as Antigravity 2.0, keeps Skills / Hooks / Subagents / Extensions (now plugins). Gemini CLI sunsets June 18. — developers.googleblog.com
• Cline SDK — @cline/sdk — Cline’s internal agent harness extracted into a standalone open-source TypeScript SDK. Same code path now powers VS Code, JetBrains, and the CLI. Plugins register tools, observe lifecycle events, add rules and commands, and shape what the agent sees. New --worktree flag auto-creates ~/.cline/worktrees/<id> so --taskId / --continue resume in isolation. — github.com/cline/cline
• OpenCode v1.2.0 — JSON → SQLite — sessions now live in ~/.local/share/opencode/opencode.db (Drizzle ORM + bun-sqlite, WAL mode). Heads-up: incremental upgrades silently skipped the migration and orphaned existing JSON sessions; some users also saw incomplete sessions in the CLI while the Desktop app kept full history. Back up the data directory before upgrading. — release
• Cursor Composer 2.5 (2026-05-18) — matches Opus 4.7 and GPT-5.5 on coding benchmarks at $0.50 / $2.50 per M tokens (standard) or $3.00 / $15.00 (fast). Built on Kimi K2.5, 25× more synthetic training tasks than Composer 2. Next-gen Cursor model in training on SpaceXAI’s Colossus 2 (~1M H100-equivalent GPUs, 10× the compute of 2.5). — post
• Codex 0.133.0 — Goals on by default (2026-05-21) — Codex’s equivalent of /goal lands enabled by default with dedicated storage and cross-turn progress tracking; goal continuations halt at usage limits rather than looping. codex remote-control becomes a foreground command with explicit start / stop and machine status. Permission profiles gain lists, inheritance, and managed requirements.toml. — release
• Simon Willison — “The last six months in LLMs in five minutes” (2026-05-19) — annotated PyCon US lightning-talk slides. The fast-forward catch-up if you’ve been ignoring the timeline. — post
• Datasette Agent (2026-05-21) — Simon’s first release of the LLM + Datasette mashup as an extensible AI assistant. — post

Next time

• TBD.

Run /news skill

Floor: 2026-05-16 (261605-1). Generated 2026-05-23.

New Claude Code commands & features

• /simplify → /code-review (v2.1.147, 2026-05-21) — renamed; takes an effort level (/code-review high), --comment posts findings as inline PR comments; the old cleanup-and-fix behaviour is removed — release
• Background session pinning (v2.1.147) — Ctrl+T in claude agents pins idle sessions so they stay alive and restart in place for updates — release
• /usage per-category breakdown (v2.1.149, 2026-05-22) — limits split across skills, subagents, plugins, per-MCP-server cost — release
• /diff keyboard navigation (v2.1.149) — detail view scrollable with arrows / j / k / PgUp / PgDn / Space / Home / End — release
• GFM task lists render (v2.1.149) — - [ ] todo / - [x] done finally render as checkboxes — release
• allowAllClaudeAiMcps (v2.1.149) — enterprise managed setting to load claude.ai cloud MCP connectors alongside managed-mcp.json — release

Claude Code — other notes

• [2026-05-21] v2.1.147 also: auto-updater retries transient network failures and reports specific error categories / OS codes; consecutive duplicate prompts no longer recorded in history — release

Codex

• [2026-05-21] Codex 0.133.0 — Goals enabled by default with dedicated storage and cross-turn progress tracking; continuations halt at usage limits instead of looping. codex remote-control becomes foreground with explicit start/stop and machine status. Permission profiles gain list APIs, inheritance, managed requirements.toml, stronger Windows sandbox integration. Plugin discovery marketplace-aware with installed versions and remote collections. Extensions can observe subagent start/stop, tool execution, turn metadata, and async approval. — release
• [2026-05-22] Codex 0.134.0-alpha train — daily alpha cuts landing on main — releases

Google I/O 2026

• [2026-05-19] “Foothills of the singularity” — Demis Hassabis frames Google’s I/O announcements as the AGI roadmap; 2030 ± 1, ~50% probability; agents + multimodal + robotics + scientific AI as the four pillars — Semafor — Fast Company
• [2026-05-19] Antigravity CLI replaces Gemini CLI — terminal-first agent spawning, same harness as Antigravity 2.0; Skills / Hooks / Subagents / Extensions carry over as Antigravity plugins. Gemini CLI + Gemini Code Assist IDE extensions stop serving June 18, 2026 — Google Developers Blog
• [2026-05-19] Antigravity 2.0 desktop app — full rebuild, parallel-agent orchestration with dynamic subagents and scheduled task automation — TechCrunch
• [2026-05-19] Gemini 3.5 Flash GA — 4× faster than Gemini 3.1 Pro, more expensive, planned as the default behind Google products — Simon’s analysis
• [2026-05-19] Gemini Omni — multimodal world model: text + image + audio + video in, physics-grounded editable video out; SynthID-watermarked; Omni Flash to AI Plus / Pro / Ultra via Gemini app and Google Flow, free on YouTube Shorts and YouTube Create — blog.google
• [2026-05-19] Managed Agents in Gemini API — single API call spins up an agent that reasons, uses tools, and runs code in an isolated Linux env with persistent state across sessions — Google I/O recap
• [2026-05-19] Native Android in AI Studio — prompt → app with integrated emulator and Play test-track publish — Google I/O recap
• [2026-05-19] Google AI Ultra ($100/mo) — 5× higher Antigravity quota plus bonus credits for overages — Google I/O recap

Adjacent tools

• [2026-05-18] Cursor Composer 2.5 — matches Opus 4.7 and GPT-5.5 on SWE-Bench Multilingual (79.8%) and CursorBench v3.1 (63.2%) at $0.50 / $2.50 per M tokens (standard); built on Kimi K2.5, 25× more synthetic training tasks than Composer 2; next-gen model in training on SpaceXAI’s Colossus 2 — post
• [2026-05-14] Cline SDK — @cline/sdk — Cline’s agent harness extracted into a standalone open-source TypeScript SDK; same code path now powers VS Code, JetBrains, and the CLI; plugin architecture with tool registration, lifecycle events, rules, commands; --worktree for isolated --taskId / --continue runs — MarkTechPost
• [2026-05] OpenCode v1.2.0 — JSON sessions → SQLite (Drizzle ORM + bun-sqlite, WAL mode); incremental upgrades silently skip the migration and orphan existing JSON sessions, so back up ~/.local/share/opencode/ before upgrading — release
• [2026-05-14] xAI Grok Build — early beta agent CLI for SuperGrok Heavy subscribers at $300/mo; 2M context, 8 parallel subagents, 16-agent Heavy architecture — Engadget

Simon says

• [2026-05-22] “The memory shortage is causing a repricing of consumer electronics” — HBM demand bleeding into smartphone RAM pricing — post
• [2026-05-21] Datasette Agent — first release of the LLM + Datasette mashup as an extensible AI assistant — post
• [2026-05-20] Google I/O coverage — Gemini Spark, Antigravity CLI transition, prompt-injection security concerns for agent products — post
• [2026-05-19] “The last six months in LLMs in five minutes” — annotated PyCon US lightning-talk slides — post
• [2026-05-19] “Gemini 3.5 Flash: more expensive, but Google plan to use it for everything” — pricing analysis — post

Anthropic — other notes

• [2026-05] Managed Agents on Claude — managed execution layer separating agent logic from runtime concerns (orchestration, sandboxing, state, credentials); long-running multi-step workflows with external tools, error recovery, session continuity — news
• [2026-05] Claude for Small Business — connectors + ready-to-run workflows across QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365

Industry

• [2026-05] Snap fired ~1,000 engineers — 65% of code now AI — controversial PR-by-PR analysis circulating; ammunition for both sides of the “agent productivity vs. fragility” debate — post
• [2026-05] Cloudflare Infire — custom AI inference engine that separates input processing from output generation across multiple GPUs; runs LLMs on Cloudflare’s global network with reduced memory usage and faster startup — InfoQ

Topics worth a 5-min slot

1. Foothills of the singularity — what was actually new at I/O vs. AGI rhetoric, and which I/O announcement are you most likely to touch in the next month?
2. Antigravity CLI replaces Gemini CLI — who’s on Gemini CLI today, and what’s the cost of the June 18 cutover?
3. Cline SDK + OpenCode SQLite — the open-source side ships meaningful architecture this week; is the Claude-alternatives gap finally closing?