lucas gu

asked Claude to check my next streetcar. it did.

@cramforce @computesdk persistent sandboxes is the big one for coding agents. the rebuild-environment-every-run pattern is the hidden cost nobody talks about. when the sandbox persists, the agent stops being a stateless tool and starts being a collaborator that picks up where it left off.

Vercel Sandboxes are now the fastest sandbox using real VMs as security boundary based on the @computesdk benchmark. The team has been absolutely cooking on this. And the best thing: Because we have a unified Fluid Compute stack across Sandbox, Builds, and Functions these wins are often shared across the stack. On the feature side there is a really exciting roadmap ahead as well. My favorites (all driven by feature requests from our customers): - Persistent sandboxes (in beta, GA immanent) - The fully mutable firewall also becomes fully programmable

@DavidKPiano AI collapsed the execution gap but widened the judgment gap. a junior with the right tools ships as fast as a mid now. what separates seniors is knowing what NOT to build. that's the part you can't prompt for.

Mid-tier devs who refuse to level up don't realize that the biggest risk isn't just AI replacing them It's also motivated junior devs easily getting to their level and then surpassing them

@DavidKPiano AI collapsed the execution gap but widened the judgment gap. a junior with the right tools ships as fast as a mid now. what separates seniors is knowing what NOT to build.

@RLanceMartin point 2 is the sleeper. we've been running persistent agents and keep finding guardrails from months ago that now actively hurt performance. building the harness is the easy part. knowing when to strip constraints as the model outgrows them is the real maintenance burden.

i co-wrote the Anthropic engineering blog on Claude Managed Agents, and wanted to share some thoughts on agent harnesses + infrastructure for long-horizon tasks ... 🧵 anthropic.com/engineering/ma…

2029063182409666572

@rjchint @katibmoe @composio @CompositeAI is cursor. @withoneai is claude code.

@katibmoe How is this different @composio

Introducing One. The simplest way to connect and monitor AI agents to hundreds of apps. And we’re open-sourcing the world’s largest integration database powering it: 47,000 agentic actions across 250+ apps. RT + comment “One” for access & 1M free API requests/month.

btw we acquired a computer use company and then 4 weeks later shipped computer use x.com/LucaWeihs/stat…

@Tailscale This is amazing. Just went to a Tailscale meetup in Toronto a few weeks ago. Tailscale works really well as a local agent and mobile relay. The only downside is that both sides need to have Tailscale installed, but there doesn't seem to be any other way around it.

API keys were never meant to live everywhere…but here we are 🔑 AI tools > agents > pipelines > containers And suddenly they do. There’s a better way! The Aperture alpha is now available via self-serve. Learn more. tailscale.com/blog/aperture-…

@tvytlx 太强了。前几周刚去了一个tailscale 在Toronto的meetup。tailscale 作为local agent和mobile端的relay也非常好用。缺点就是两边都要装 tailscale，但好像也没有其他的办法。

Tailscale 昨天发布了 Aperture，想把 AI 场景里的“API key 到处散落”问题，收敛成一个统一的 AI 网关问题来解决。 Aperture 本质上是一个 AI gateway + governance layer。 我们经常碰到 API key 在 .env 里一份，CI 里一份，Docker 里一份，每个开发者本地一份，agent runtime 又一份。哪个 key 还在用、谁在花钱、调了什么模型，很难搞清。 Aperture 主要功能有： - 集中管理 LLM API keys：每个 provider 只需要把 key 放在网关上，不再下发到每个开发者、CI、容器或 agent。 - 基于 Tailscale 身份鉴权：请求不是靠分发 key，而是靠 Tailscale identity 识别是谁、哪个服务、哪个 agent 在调用。 - 统一接入多种模型：支持 OpenAI、Anthropic、Google，以及自托管/私有 OSS 模型。 - 兼容主流 coding agents：支持像 Claude Code、Codex、Gemini CLI 这类可以自定义 base URL 的 agent，也支持相关 agent frameworks。 - 完整会话与用量可观测：可以看 LLM session history、token 使用、模型使用、tool calls，便于发现异常、识别高价值工作流、看成本分布。 - MCP / 本地工具调用提取与控制：不仅看模型请求还能看到 MCP 和本地工具调用；官方页面写到已经支，持提取这些调用，并且支持更细粒度控制。 - 访问策略与模型授权：配置里可以定义哪些用户/服务能访问哪些 provider / model。 - Hook / 外部集成：可以按条件触发 webhook，把特定请求或 tool call 发给外部系统。 - 日志导出：当前支持导出到 S3。 - 适合沙箱和 CI 场景：官方明确提到可用于 GitHub Actions 这类容器化/沙箱环境，只要能跑 Tailscale。 x.com/Tailscale/stat…

Use /schedule to create recurring cloud-based jobs for Claude, directly from the terminal. We use these internally to automatically resolve CI failures, push doc updates, and generally power automations that you want to exists beyond a closed laptop

@jakexia27 @kevinyang damn!

there's so much performative vibe coded agent slop. anyone can vibe code an agent but wrangling it to really work for you is the hard part. super excited for @kevinyang to help people actually get shit done!!!

@yetone local first 100%

Alma 还是压中了 local-first 时代

@zarazhangrui I want "bring your own subscription" more than BYOK if anthropic keeps claude max plan forever.

Trend I'm seeing: instead of paying for various model wrapper products each charging a subscription, I'm just gonna own model API subscriptions and then use various agent skills to get a more personalized version of the wrappers "Bring your own API key" will become a thing

@yetone 完全同意。这个和当年notion 搞template一样。我统称为，onlyfans for personal productivity.

这些花活儿我称之为 Agent Porn，我大学就已经过完 Unix Porn 的瘾了，现在已经贤者模式了 更何况很多 Agent 功能都应该是 Agentic 出来的，不应该是 human aware 的东西

@dani_avila7 /rename will automatically apply the session name and color.

Claude Code now lets you customize the prompt bar color Mine is orange, and it finally matches my Ghostty config and desktop wallpaper Small thing, but it matters when you stare at your terminal all day 🤩

@dani_avila7 i love this.

@OmidMogasemi the .claude folder is still not backed up and synced natively. i made a temporary solution before it is officially supported. github.com/lucasygu/claud…

You can now add a setup script in Claude Code on the web & desktop! Use these to automate setup before Claude Code launches on a cloud environment. It’s particularly useful for installing dependencies, settings, configs, etc.

@luoling8192 哈？

我草，一觉醒来 Github 被封号了

Investigating: Elevated errors on @claudeai stspg.io/xs6ftr1kgydn This is affecting the login/logout actions in Claude Code as well.