Father, founder, meat proxy.
Austin
Joined July 2009
- Tweets705
- Following42
- Followers125
- Likes94
Pinned Tweet
agents don't belong on your laptop.
wrote the 15-min setup for giving every task its own throwaway cloud box: copy-paste commands, safe-secrets rules, and a prompt to paste to your agent.
my agents kept telling me "done" when nothing was done.
so i changed what done means:
- a pushed branch
- the exact verify commands in AGENTS.md pass
- CI green after review
- no --no-verify, ever (hooks block it)
now "done" is something i can check, not a vibe. what's your definition of done for agents?
jev-latest flips to jev 1.14 around oct 21. tuned confidence thresholds? pin jev-1.13.0 and run your evals first. full guide: what changed, images + box + 262k context (danger zone), 7 business use cases, and the tricks the docs actually back up
3 questions i keep getting about running agents in throwaway cloud boxes:
"what if the timer kills unpushed work?"
→ agent pushes a branch + opens a draft pr before it says done. the box never holds the only copy
"doesn't a fresh box mean cold context?"
→ context lives in the repo. AGENTS.md with exact verify commands + a snapshot with deps installed. fresh box, warm start
"doesn't this just move the risk to your secrets?"
→ one fine-grained token per task, one repo, short expiry. the vm only sees a placeholder
what would stop you from moving your agents off your laptop?
Going to give this a try and report back - my guess there is a lot of wasted tokens and time hiding in the shadows and too much to notice or review without tooling like this
Introducing AgentPlay 🚀
The future of coding is self improving coding agent harnesses that help agents avoid making the same mistakes over and over.
Turn your existing Claude Code, Codex, Cursor and Copilot sessions into interactive maps and replayable timelines.
See what your agent explored, where it failed, what it retried and how it reached the final change
Free, local and open source.
GitHub: github.com/Entelligence-AI/a…
things that actually made my coding agents faster this month, with numbers:
- found 52 stale agent worktrees eating 51gb on my mac mini. cleaned to 5.6gb
- gave agents a faster file search: median "find this" task went 43.3s → 10.2s across 94 repos / 210gb
- but with no rule, agents still picked the built-in tool 12/16 runs
- one line in AGENTS.md → 16/16. total tool time 715s → 7s
- "a missing check script means the command failed." agents love to call a missing script a pass
- --no-verify is banned
the one-line rule did more than the tool. what's in your AGENTS.md that actually changed behavior?
if you're on MacOS your coding agents spend most of their time just looking for files. here's how to fix it in one command
benchmarked on my dev folder (94 repos, 210GB):
- find a file by name: 2.4s → 6ms
- files containing an identifier: 15.3s → 4ms
- where a function is defined: 15.0s → 3ms
- agent's median "find the thing" task: 43s → 10s
same answers, same token cost. it was all time spent waiting on tools
the trick is fsearch by Noah Dunnagan, an indexed whole-disk search
the part nobody tells you: installing it isn't enough. agents still used their built-in Glob/find 12 of 16 runs.
i packaged it as a skill for claude code, codex, cursor and opencode. link below 👇
skill + full benchmark method + a script to reproduce it on your machine:
github.com/gregkop/agent-fse…
all credit for the search engine to github.com/noahdunnagan/fsea…, i just wired it into agents and measured it
cheap trick for agent code review:
don't let the reviewer see the chat
give it only the diff + the spec and tell it to find what breaks
if it reads the whole conversation it just agrees with the author
came across this and wanted to test it, pointed it at my local Obsidian instance with a prompt to reverse engineer how Obsidian builds its internal knowledge graph, it was able to do it and create this video explaining how it works
Someone gave Claude Code a decompiler. 😂
Software engineering is getting ridiculous.
There's an open-source project called REA (Reverse Engineer Anything) that lets coding agents investigate applications without their original source code.
It connects agents to Ghidra, Hopper and IDA through MCP. They can inspect binaries, read decompiled functions, trace calls and investigate how software works internally.
And the examples are wild.
The project traced how Notion handles clipboard operations through Electron IPC. It also reconstructed a function from the 1996 game DX-Ball, verifying it against 3,205 original-x86 test cases.
Of course, this still requires analysis and validation. It isn't magic source-code recovery.
But we've gone from asking AI to write software to asking AI to reverse engineer software.
Next sprint: reverse engineer whatever the coding agent shipped last sprint. 😂
GitHub: github.com/morluto/rea
so my mac mini kept running out of disk space and I couldn't figure out why
turns out my AI coding agents had left 52 git worktrees lying around. 51 GB of node_modules and build output from tasks that were done weeks ago
turned the cleanup into an agent skill. it does a dry run first and never touches uncommitted work
51 GB → 5.6 GB. how many worktrees are sitting on your machine right now?
open source, works with claude code, codex, cursor: github.com/gregkop/worktree-…
so this is interesting, was doing some ad campaign research and asked OpenAI's Dot to help me research competitor ads on Facebook Ad Library and it couldn't - asked Muse and it got right after it, emailed me a zip file of all the content
agree, and its not as hard as you may think - there are OSS versions like openbot.run/ or if you want to get more in the weeds you can combine @herdrdev and @boxd_sh for similar fleet of agents each with their own computer type experience
that way you are not constrained to any single model provider, usage plan, or even to hosted models if you like running local
Oh wow, I better go install this right away
Antigravity 2.0 now features a dedicated planning mode, just like the Antigravity CLI.
Type /plan and the agent steps back to think through the task, and conduct extensive research, before generating an implementation plan for your review. The agent will ask for your approval before it begins executing on the task.
You can also ask for a plan naturally in your prompt, for a lighter version of a plan.
Jev is quite good at quickly scanning @posthog session recordings and flagging ones worth exploring further
curious if anybody has heard on their connector applications yet? wondering how long it takes for review
Opening access for developers to build Muse connectors. You bring the API -- Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there.
New connectors are live today. Come build with us. muse.ai/platform
Ran Qwen-Image-2.1 locally on an RTX 3060 (12GB VRAM), Ryzen 5 2600X + 16GB RAM. Q4_K GGUF via stable-diffusion.cpp/Vulkan with disk offload. YouTube thumbnail from a reference image: 1024×576, 40 steps, 5m 55s. Goal was to generate a thumbnail for this youtube video youtube.com/watch?v=3wZ7NhUY…
not bad at all for a local model
Meet Qwen-Image-2.1, the most balanced and cost-effective image generation model in the Qwen-Image series! Now open weights! 🎨
A unified model for both generation and editing, delivering top-tier quality in a lightweight package.
Highlights: 👀
- Compact & exceptionally fast: A lightweight 7B architecture that outperforms most closed-source models, with drastically accelerated inference for multi-image inputs.
- Native transparency: Natively generates and edits RGBA layers, enabling seamless compositing and text editing within transparent images.
- Versatile, high-fidelity editing: Supports up to 10 reference images and precise local control while preserving strict fidelity for portraits and products.
- Broad coverage & stunning aesthetics: Excels at panoramas, infographics, and virtual try-ons, delivering realistic textures and elegant typography.
Start to create your next masterpiece with Qwen-Image-2.1! 🖼️
- Blog: qwen.ai/blog?id=qwen-image-2…
- GitHub: github.com/QwenLM/Qwen-Image…
- Model Scope: modelscope.cn/models/Qwen/Qw…
- Hugging Face: huggingface.co/Qwen/Qwen-Ima…
