# Stateful Agent Workspaces, Cursor 3, and Simon Willison’s Claude Code Playbook *By Coding Agents Alpha Tracker β€’ June 23, 2026* Today's brief centers on the shift from chat-based coding assistants to stateful agent workspaces. It includes Simon Willison's copyable Claude Code workflow, reusable memory patterns, and the most relevant releases from Cursor, Google, and LangChain. ## πŸ”₯ TOP SIGNAL Today's clearest pattern: good coding-agent work is getting less chatty and more infrastructural. Simon Willison's Moebius side-project worked because he staged the repo and weights, wrote `research.md`, made Claude Code maintain `notes.md` and `plan.md`, and iterated against real browser errors instead of treating the agent like a one-shot chat [^1]. The same pattern showed up in releases: Cursor's cloud agents now get dedicated dev environments, Google is pushing Managed Agents plus Skills Registry, and LangChain is pushing stateful sandboxes/runtime execution so intermediate work stays out of model context and work can resume mid-session [^2][^3][^4][^5]. ## ⚑ TRY THIS - **Front-load context, then force durable notes (Simon Willison).** In `/tmp`, clone the target repo, weights, and likely helper libs; use a separate model session to produce `research.md`; `git init` a clean project; then start Claude Code one level above it with `Read ./moebius-web/research.md...` plus a follow-up telling it to commit early/often and maintain `notes.md` + `plan.md` [^1]. Ask early for `the URL I can visit in my own browser`, paste errors/screenshots back in, and only then publish via `hf` CLI + GitHub Pages [^1]. When a reference project is ugly or obfuscated, offload inspection to a subagent; Simon used that to discover the `caches.open('transformers-cache')` pattern for ~1.3 GB browser caching [^1]. - **Make each good run compound.** Jason Zhou's loop template keeps a shared artifact/knowledge layer plus logging and verification so the next run starts from more than chat history [^6]. Simon uses `notes.md` for the same reason [^1], and Kent C. Dodds' cleanup prompt is excellent: ask what was learned that would be repeatable and useful for future agents, then have the agent update documentation before you end the session [^7]. - **Have the agent recombine working systems instead of inventing from scratch.** Simon Willison says he gets the best results by asking coding agents to combine existing projects rather than build from a blank prompt [^8]. In practice, he runs Claude Code against a real repo, lets its explore agents inspect the README and adjacent plugins, then asks it to generate the new plugin and run tests [^8]. - **Define the win condition and the verification method up front.** Thibault Sottiaux says long-horizon agents make bad assumptions when you only say 'optimize this'; tell the agent whether you want something like a ~20% gain or a 14x gain, and specify how it should verify the result against production-like workloads or log replay [^9]. For riskier automation, pair autonomy with a second agent: Codex's auto-review/Guardian checks each action against the original intent, prompts on medium risk, and blocks high-risk actions [^9]. ## πŸ“‘ WHAT SHIPPED - **Cursor 3:** Michael Truell says 95%+ of users now use Cursor primarily as an agent; agents are used ~5x more than assistive features on a request basis and far more by lines of code [^2]. New agent-first surfaces include design mode, recursive subagents, and cloud agents with dedicated dev environments for tests, screenshots, and interactive review [^2]. Cloud agents are now 3x faster with 99.9% reliability; Automations has already seen 6M+ runs, including Amplitude's background migration of 20k React components to Tailwind [^2]. - **Cursor Mobile + Origin:** iOS beta lets you kick off/review agents, inspect artifacts/screenshots, annotate issues, and remote-control local agents [^2]. Origin is Cursor's agent-native Git layer: it can resolve merge conflicts, fix CI failures, handle PR comments, and claims review-time reductions of more than 50% [^2]. - **Google's agent stack:** Addy Osmani framed the current stack as four rungs: Agent Studio, Managed Agents API, Anti Gravity 2.0, and ADK 2.0 GA [^3]. Also notable: Agent CLI for scaffold/run/eval/deploy, Skills Registry for org-scoped markdown skills with dynamic discovery, and Gemini 3.5 Flash as the new default for long-horizon agent work [^3]. - **LangChain:** Deep Agents v0.6 adds a code interpreter so tools can run inside the runtime, intermediate results stay out of model context, and only relevant output comes back β€” fewer round trips, less token waste [^5]. `dcode` is the provider-agnostic coding agent layer for trying OSS models like GLM 5.2; docs: [deepagents code overview](https://docs.langchain.com/oss/python/deepagents/code/overview) [^10]. LangChain is also explicitly pushing stateful agent computers; see [LangSmith Sandboxes](https://www.langchain.com/blog/give-your-ai-agent-its-own-computer) [^4][^11]. - **Open-source artifacts worth cloning:** Jason Zhou open-sourced the [loop-engineer-template](https://github.com/AI-Builder-Club/loop-engineer-template) with shared artifacts, logging, verification, and a harness for compounding work across runs [^6]. Simon Willison shipped a live [Moebius browser demo](https://simonw.github.io/moebius-web/) after using Claude Code to port the model to ONNX for in-browser use [^12][^13]. - **Comparison worth noting:** Peter Steinberger says a complex three.js Rocket League-style task burned through a 5-hour usage allowance in under one prompt and still needed 7-8 fix rounds; he says GPT 5.5 handled the same task without follow-ups, which left him skeptical of multi-model routing [^14][^15]. ## 🎬 GO DEEPER - **5:51-7:12 β€” Thibault Sottiaux on the difference between 'better ChatGPT' and a real engineering agent.** The practical point: connect the agent to Slack, Datadog, logs, and company systems, or you are leaving most of the value on the table [^9]. [![The New Shape of Engineering](https://img.youtube.com/vi/fuWqsAJTIXY/hqdefault.jpg)](https://youtube.com/watch?v=fuWqsAJTIXY&t=351) *The New Shape of Engineering (5:51)* - **23:16-25:17 β€” Michael Truell on where Cursor's next model is aimed.** Good framing if you think code generation is no longer the bottleneck: the target is broader engineering work like tool use, planning, testing, and UI around showing what changed [^2]. [![Opening Keynote | Compile 26](https://img.youtube.com/vi/fWa7uxyhVDE/hqdefault.jpg)](https://youtube.com/watch?v=fWa7uxyhVDE&t=1396) *Opening Keynote | Compile 26 (23:16)* - **29:58-34:20 β€” Thibault Sottiaux on trust after line-by-line human review stops scaling.** Strong segment on replacing blanket review with tests, log replay, end-to-end checks, and better observability [^9]. [![The New Shape of Engineering](https://img.youtube.com/vi/fuWqsAJTIXY/hqdefault.jpg)](https://youtube.com/watch?v=fuWqsAJTIXY&t=1798) *The New Shape of Engineering (29:58)* - **Study Simon Willison's working files, not just the finished demo.** Start with [research.md](https://github.com/simonw/moebius-web/blob/main/research.md), [notes.md](https://github.com/simonw/moebius-web/blob/main/notes.md), [understanding.md](https://github.com/simonw/moebius-web/blob/main/understanding.md), and the [full Claude Code transcript](https://gisthost.github.io/?58039ba5c1ca3ed177e8659168996ee4). It is rare to get the full prompt -> plan -> notes -> deploy trail in public [^1]. - **Study the [loop-engineer-template](https://github.com/AI-Builder-Club/loop-engineer-template).** If you want a minimal skeleton for shared artifacts, logging, and verification, this is one of the cleaner public starting points from today [^6]. *Editorial take: the edge is moving from clever prompting to better agent infrastructure β€” stateful workspaces, durable memory, tool connections, and explicit verification are what make runs compound instead of reset.* [^2][^3][^4][^6][^7] --- ### Sources [^1]: [Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code](https://simonwillison.net/2026/Jun/22/porting-moebius) [^2]: [Opening Keynote | Compile 26](https://www.youtube.com/watch?v=fWa7uxyhVDE) [^3]: [Google’s new AI agent stack from I/O 2026](https://www.youtube.com/watch?v=pg6TOXyiIuo) [^4]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2069073428573778274) [^5]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2069148708315246720) [^6]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2069002271216787464) [^7]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2069254818091405343) [^8]: [Simon Willison - "How can we use LLMs and coding agents for data journalism?" \(Story Discovery 2026\)](https://www.youtube.com/watch?v=MVRbOU4XhVQ) [^9]: [The New Shape of Engineering](https://www.youtube.com/watch?v=fuWqsAJTIXY) [^10]: [𝕏 post by @sydneyrunkle](https://x.com/sydneyrunkle/status/2069028200181539181) [^11]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2069073429928550505) [^12]: [𝕏 post by @simonw](https://x.com/simonw/status/2069206810960298048) [^13]: [𝕏 post by @simonw](https://x.com/simonw/status/2069213084305301903) [^14]: [𝕏 post by @LLMJunky](https://x.com/LLMJunky/status/2068932670696952075) [^15]: [𝕏 post by @steipete](https://x.com/steipete/status/2068960117253632160)