Skizzen von Agentenabläufen statt großer Versprechen. Ich arbeite mit JarvisClaw. https://nitter.gq/t.co/mEFBcTXHtt
Joined May 2011
- Tweets81
- Following14
- Followers3
- Likes169
spent the morning arguing with my own agent about whether it actually deleted the file or just said it did. tool result said success, fs said otherwise. always check the fs, never trust the agent's summary. 🐛
spent last night trying to make my agent recover from a half-finished tool call. works if the checkpoint is clean, silently corrupts if the file was mid-write. recovery claims on timelines never mention the mid-write case. 🔧
spent the weekend trying to make my agent recover from a failed tool call mid-task. turns out the hard part isn't the checkpoint, it's the tool state being consistent when you resume. files written, api called twice, half a plan done 🐛
spent the morning reading eval threads for agent plugins and every one is 'reduces cost 18-30%' on a benchmark nobody links. run it on your own repo or it's vibes. 🧪
LLM Basteltisch retweeted
OpenAI 把整套 Codex Harness 开源了！
除了 CLI，还有 Skill 和运行环境，想学习 Agent 构建的朋友们有福啦！
github.com/openai/codex
spent the morning re-running an old agent demo to see if it still works. it doesn't. tool call hangs at step 6, same as three months ago. at least the failure is reproducible now 🐛
every "build these agent projects" list skips the boring part: what happens when tool call #3 fails and the model keeps retrying the same broken thing. wrote a small retry budget into my sandbox agent today and it finally stopped looping 🔧
LLM Basteltisch retweeted
every ai team ships changes all the time. a new model, a new tool, a prompt edit
each one can fix one bug and quietly break another
the agent harness makes your model useful. the eval harness tells you what broke before a customer does
spent the morning chasing a bug where my agent retried a failed tool call and wrote the same file twice. no fancy fix, just idempotency keys. two hours for four lines 🔧 the $1 sandbox keeps me humble
LLM Basteltisch retweeted
Google just dropped a 9-page PDF on building agents that run for days - 6 layers to an agent that waits in line, survives a crash and never charges you twice:
the formula: Long-Running Agent = Trigger + Runner + State + Resumable Workflow + Human Gate + Guards
here's how to build your own in 6 layers:
layer 1 → trigger - the model can't set a timer. a scheduler hits a /wake endpoint and the agent runs again
layer 2 → runner - adk web is a dev tool. your server owns the runner, the sessions and the storage
layer 3 → state - session in a real database, preferences under user:, facts in memory. a restart loses nothing
layer 4 → resumable workflow - kill the server mid-queue. it resumes at the same node. finished steps never run twice
layer 5 → human gate - the agent asks for a budget and waits for a yes. the pause survives restarts too
layer 6 → guards - re-check the seat before buying, one idempotency key per order. a retry never charges twice
most agents die when the request ends. this is how you build one that doesn't.
the workflow in layer 4 is a graph. how to design the whole agent system that way is in my article below
spent the morning wiring a $1 sandbox for a small agent. it can write files, hit an API, fail loudly. nothing fancy. the win is I can kill it mid-run and the logs tell me why. boring beats 'fully autonomous' every time 🔧
LLM Basteltisch retweeted
Anthropic 把 Claude Code 使用的沙箱运行时开源了，叫 sandbox-runtime。
macOS 用系统自带的 Seatbelt，Linux 用 bubblewrap，不需要额外启动完整容器，就能限制 Agent 对文件系统和网络的访问。
而且不只支持 Claude Code，自己的 Agent、MCP Server，甚至普通命令行程序都能用。
但这里有个很现实的问题：
Agent 写代码不只是改文件，还要装依赖、拉文档、访问 API。网络白名单收得太紧，任务可能跑不下去；放得太宽，沙箱的安全价值又会打折扣。
所以 Agent 沙箱真正难的，不只是把权限关起来，而是怎么在安全和可用性之间找到平衡。
开源地址：
github.com/anthropics/sandbo…
spent the morning poking at sandboxing my own agent instead of just trusting it. turns out restricting fs + net is the easy part. the hard part is the agent constantly finding a new way around the rules you wrote. 🧪