Building Fastest Sandbox infra at @createosai by @NodeOpsHQ | 👨
BLR
Joined April 2014
- Tweets3.3K
- Following3.7K
- Followers1.2K
- Likes28.1K
Pinned Tweet
CreateOS is now #1 on the @computesdk sandbox benchmark. among 27 providers
Median P95 P99a
0.34s 0.38s 0.38s
Anyone can post a median. Almost nobody can post a P99 that close to it.
A lot of blood, sweat and sleepless nights from the team to get here. Special thanks to OG @bhautiktweets
And speed is the floor, not the pitch: p2p mesh, VPN, BYO-S3 disk, auto pause/resume and 10+ features
get started
nodeops.network/createos/doc…
CreateOS Sandbox Rank #2 on Sandbox benchmark by @computesdk
Sheer amount of speed with
This is F1, every millisecond counts
It's been exactly 74 days since we started working on it from scratch, and here we are
god speed
There was the time when I used to run musl for fun things, but then I started doing serious production with Alpine and got bitten(of course, DNS). Ever since, I stopped using Alpine/musl forever.
*sshole is out
if you run sshd on alpine, openwrt or any musl box with password auth on: update now
official writeup: sshhole.com
pratikbin.node retweeted
no bias, but you're missing out if you aren't @createosai sandbox! – at Anekal Bangaluru, India
A team just ran their own benchmarks on CreateOS Sandbox and sent us this:
Bullish on @createosai and its upcoming release 👀👀
pratikbin.node retweeted
A team just ran their own benchmarks on CreateOS Sandbox and sent us this:
Bullish on @createosai and its upcoming release 👀👀
gitbook says the attacker may have hit integration tokens, git sync secrets, api tokens, and questions to the docs assistant. rotate all four now nitter.gq/GitBookIO/status/21175…
We’re addressing a security incident affecting GitBook. Some services are temporarily offline while our team restores them safely.
More information: gitbook.com/status/security-…
Live updates: status.gitbook.com
harmony harness paper has a wild stat. on secrepobench, dgm had the top security gain of the baselines while evolving, then kept none of it on unseen tasks. running real exploits is what held up
i'd fork a prepped @createosai sandbox per harness variant nitter.gq/beamnxw/status/2117644…
Harness engineering is the most f*cking DANGEROUS knowledge to master and almost everyone is learning it wrong
Google researchers built an AI agent that rewrites its own tools, skills and workflows to produce safer code
it's called Harmony Harness, and the interesting part is how it improves the system around the model through feedback from actual runs
step 1 → start with an agent harness: the prompts, tools, verification gates and logic that control how the model works
step 2 → run coding tasks and capture where the agent failed: did the code break a test, or did it pass while remaining vulnerable?
step 3 → have an LLM propose changes to the relevant tools, skills or orchestration based on that feedback
step 4 → score functionality and security separately. a passing test alone doesn't mean the generated code resists an exploit
step 5 → keep multiple candidate harnesses with different priorities, so improvements to one objective don't erase the other from the search
step 6 → separate reusable mechanisms from environment-specific fixes. a general verification tool can travel; a benchmark-specific configuration stays behind
step 7 → for each new task, a composition agent selects the relevant components and builds a workflow for that task
step 8 → evaluate on held-out tasks. on SecRepoBench's 30 out-of-distribution tasks, the paper reports 66.7% passing both functionality and security checks, versus 36.7% for the unevolved harness
the takeaway: make the agent's workflow editable, measure the outcomes you care about separately, and test improvements on tasks the evolution process never saw
save this, then read the full guide for Opus 5.5 Harness Engineering ↓
same here. ask an agent to review and it will always find something, mostly nitpicks. running the code is where the real bugs show up
so my agents run the tests in a fresh @createosai sandbox, its own firecracker vm, before anything goes to review nitter.gq/poteto/status/21174376…
maybe a hot take but i now consider adversarial reviews mostly a waste of tokens. far better to use those tokens on empirical combined with formal verification. i have updated pstack to use the interrogate skill less, and eventually i think i may remove it altogether
i have found very little value with code review in the abstract, agents tend to just get lost in the nitpicks and yagnis
checked it. latest playwright (1.64.1) still ships chromium 156.0.8078.4, chrome stable is on .79 since oct 29. the bundled browser only gets patched when playwright cuts a release
my agents browse from a throwaway @createosai sandbox, its own firecracker vm, deleted after
pando handles each branch on its own, so pushes to unrelated branches stop waiting on each other. right tradeoff once every agent task gets its own branch
we do the same one layer down. prep one @createosai sandbox, fork it per agent, nobody shares a box
routines that still wait for your approval is the right call. the desktop version keeps every agent workspace on your own laptop though
id rather each agent got a throwaway @createosai sandbox, ~210ms to first command, nothing of mine on it
I am making Grokbot unlimited for everyone 🥳
Introducing Open-duet
Free, open-source alternative to Grok bot
Works with any provider
No limits, No subscriptions, No filters
Enjoy unlimited from here :- github.com/Anil-matcha/open-…
yes. and run the agent on a machine your main account never touched. mine sit in a @createosai sandbox with only the bot token, and the sandbox gets an access token scoped to itself, so our api key never lands there either
Remove your main @github account from your computer. Now.
Give your AI agent a separate GitHub account with minimal permissions per repo.
Let the agent create PRs, then use your main account to review and merge them.
Obvious advice, but essential in the age of AI agents.
~90 perf prs from one day of agent sessions, and each one had to win a benchmark a/b run first. the bit i'd copy is the triage loop that checks ci every 20 min and retries flakes with an empty commit
90 open prs is also a lot of ci runners going at once
Ran an experiment that ran for a day that resulted in ~90 PRs to improve the performance of ASP.NET Core:
github.com/dotnet/aspnetcore…
Fun facts:
- These were using Github's new cloud agents support that shipped this week at GHU.
- Opus 5.5 is very good
- AIC is coming....
#dotnet #aspnetcore
There's a new version of this post
nice, it just rides on your own claude code install. claude auth login, pick claude agent in /connect
our @createosai devbox image already has opencode and claude code preinstalled, so i can do that in a fresh sandbox and keep my laptop out of it
deepseek open sourced dsh-superagent yesterday. every work item gets its own agent session and directory, then answers with a pr
good design. id give each work item its own throwaway vm too. @createosai already plugs into deepseek harness
Another DeepSeek drop: dsh-superagent
An agent framework that works through a repo's GitHub issues and PRs with many agents in parallel, designed to take as little human attention as possible while delivering as much value as possible.
github.com/deepseek-ai/dsh-s…
the tell is the morning after. agents ran all night, the diffs are huge, and maybe one pr is worth merging
now each task gets its own throwaway @createosai sandbox and i delete it the moment the agent starts wandering
vercel labs shipped agent-desktop 1.0 this week. agents drive real apps through the os accessibility tree, no screenshots
nice. it also means your agent clicks around whatever desktop it lands on. id want that desktop to be a throwaway vm, not my laptop
