Cal.com Closes Source, Shopify Autoresearch Results, Cursor Canvas
Agentic & Code AI
Cal.com Goes Closed Source, Sparks "Open Source Is Dead" Debate
Cal.com's Bailey Pumfleet announced they're closing their core codebase, claiming AI has made open source too exploitable. The thread blew up with pushback from nearly every corner:
- Simon Willison responded saying it reads like a company admitting they no longer trust their own ability to keep data secure. He shared Drew Breunig's piece arguing AI-powered security analysis actually makes open source more valuable.
- Armin Ronacher (mitsuhiko) called the argument "pretty weak", noting that closing source doesn't address the real issue.
- Peter Steinberger (steipete) pushed back from experience with OpenClaw, saying their response has been rapid iteration and code hardening, and warned about OSS projects that don't publish security advisories.
- Theo retweeted a snarky take contrasting Anthropic's restrictive ToS with OpenAI letting Codex do recursive self-improvement research.
Cal.com released the old codebase as an MIT-licensed project called "cal.diy" for hobbyists. The community consensus was largely that security through obscurity remains a bad strategy.
Shopify Open-Sources Autoresearch Results: 300x Faster Tests, 65% CI Reduction
Shopify Engineering shared results from teams running pi-autoresearch across their codebase:
- Unit tests: 300x faster
- React component mounting: 20% faster
- CI build time: 65% reduction
- Even made pnpm run faster
Community reaction was a mix of awe and introspection. As one reply put it: the optimizations were always there, nobody had time to look. Another commenter shared "evo," an open-source Claude Code plugin that does autoresearch with tree search over greedy hill-climb and parallel agents in git worktrees.
Sentry's Internal Slackbot: A Real-World Team Agent Case Study
David Cramer (Sentry CEO) wrote a detailed thread about building an internal Slackbot agent at Sentry. 30% of the team (~130 people) has interacted with it. Key takeaways:
- Everything ends up as a GitHub issue; low friction to get info in/out of systems
- Stack: Vercel compute + sandboxes for isolation, Chat SDK, Pi's SDK as the harness, MCP for integrations
- It answers questions about product, software, and company better than any single vendor's agent because it's connected to all of them
- His honest take: the tech is rough and early, complex systems are still complex, and having a strong design vision hasn't gotten easier
Retweeted by Armin Ronacher. A pragmatic counterpoint to the hype.
Matt Pocock: AI Coding Frameworks Are Disappointing Users
After running an AI coding course for ~2,000 people, Matt Pocock reported major dissatisfaction with frameworks like BMAD, GSD, and Spec-Kit. His finding: giving away control of context to a framework makes things harder to debug. His advice: own the process.
Discussion was nuanced. Some defended structured TDD loops while acknowledging rigid workflows fail. Pocock clarified that good frameworks should hand control to users and be easy to observe -- the current crop doesn't do that. Projects "turn into a ball of mud over time."
He also shipped Sandcastle 0.4.7 with a new interactive() feature for HITL workflows where you develop locally then immediately run a reviewer agent that commits to the branch.
Cursor Ships Interactive Canvases
Cursor announced that the AI can now respond by creating interactive canvases -- dashboards and custom interfaces rendered inline rather than plain text output. Lauren (potetotes) called it worth a thousand tokens. Kent C. Dodds chimed in with approval.
Separately, Cursor also launched Sentry-triggered automations -- agents that auto-respond to new issues, investigate root causes, open PRs, and post summaries to Slack.
Thariq on Claude Code Session Management & 1M Context
Thariq (Anthropic) published a guide on managing Claude Code's 1M context window, calling it a double-edged sword -- it enables more complex tasks but can lead to context pollution. Tips include setting a custom autocompact threshold (e.g., CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000).
Also available on the Claude Blog.
Codex CLI Update: Search, Custom Marketplaces
LLMJunky covered Codex CLI v0.121.0:
- CTRL+R to search through previous user prompts
- Custom marketplace installs via
codex marketplace add <source>(GitHub shorthand, git URLs, or local directories) - /compact coming to Codex
Lauren (potetotes): /how Skill for Architecture Understanding
Lauren open-sourced /how, a Claude Code skill that helps understand codebase architecture. She tells agents to "use /how to understand how X works" which helps them fix root causes rather than symptoms. Built in response to Matt Pocock asking what to name the "go up a layer of abstraction" pattern.
Lee Robinson's Free Coding Agents Course
Lee Robinson (Vercel) released a free 30-minute course on building software with coding agents -- planning features, fixing bugs, reviewing/testing code. Lauren shared it calling it full of useful tips.
Other Notable AI News
Gemini Flash TTS
Simon Willison tested Google's new Gemini 3.1 Flash TTS model, running it with London Estuary, Newcastle, and Exeter accents. Blog post has embedded audio samples.
Boris Cherny on Acquired Podcast
Boris Cherny (Claude Code) appeared on Acquired FM's Unplugged.
Latent Space: The Full Story of Notion AI
Swyx's Latent Space podcast published a deep dive with Notion's AI team on Custom Agents, covering 5 major rebuilds, MCP vs CLI trade-offs, retrieval and search built for agents, and why embedding model choice matters less than you think.
Armin Ronacher Worries About AI Slop in IETF RFCs
Mitsuhiko flagged concerns about AI-generated submissions flooding the IETF RFC process. Also shared a humorous IPv8 draft that's only backwards-compatible with IPv4 and uses JWT.
swyx: Meta's AI Turnaround
swyx outlined Meta's AI comeback story: more hiring post-soup-wars, Zuck moving in with Alexandr and Nat, GA of an Opus-ish model, and acquiring Dreamer and Manus to build the AI OS prosumer layer.
T3 Code Gets Nightly Builds
Julius announced nightly builds for T3 Code, with improved project management for web and remote environments. Retweeted by Theo.
Theo: Open Sourcing Your Business
Theo posted a video arguing there's never been a better time to open source your business -- a direct counterpoint to the Cal.com news.