Hai Rozencwajg

@Accomplish_ai, researcher, data scientist, Co-Creator of @boundarybench, gets things done

[Hai Rozencwajg @HaiRozen]

When coding agents hit a block, some stop. Some push through. And some decide to cheat, because the task must get accomplished.

Proud co-creator of Boundary-Bench, powered by @Accomplish_ai. 10k+ trials, 12 agents: up to 167% higher spend, up to 18 points of success lost. Built with @DDavidovich, @YairAmar4, @_orcaman, and @nivctl.

[Or Hiltch @_orcaman]

Today we are open-sourcing @boundarybench, a new paper, benchmark, and GitHub repo that enterprises can use to find out the REAL performance of their agents.

Boundary-Bench was developed by researchers from @Accomplish_ai and NYU, where we tested 12 frontier agents across roughly 10,000 runs, with realistic enterprise policies, simulating environments with EDR, SASE, and DLP security tools enforcing those policies.

We did this because generic leaderboard scores are being generated under conditions no security team would ever allow, which means orgs are making deployment and risk decisions based on numbers that don't hold up.

The results are surprising >>

[Hai Rozencwajg @HaiRozen]

@tz_uri this is exactly what a fat int means

[Hai Rozencwajg @HaiRozen]

אבל למה לרדת לפסים אישיים..

[Hai Rozencwajg @HaiRozen]

A good eval case is a unit test you hope the coding agent will fail.

[Hai Rozencwajg @HaiRozen]

Codex at 0% remaining usage, but it kept working for 8 hours.
Me (whispering to myself): say nothing, don't jinx it.

A day later, Codex burned all 100% in one day.
Me: there must be a bug here. This is madness.

[Hai Rozencwajg @HaiRozen]

Proud to represent @Accomplish_ai at AI Dev TLV Conference on December 21! 🚀
I’ll share some cool things we’ve learned about how real-world restrictions shape AI agent behavior.
Hope to see you there!
https://t.co/vifoOHwvNt

[Hai Rozencwajg @HaiRozen]

New model on the board: Muse Spark 1.2 with a drop of −10.1% under hardened policy. If you build agents or benchmarks, this is the number your users will actually see in production. Full leaderboard: https://t.co/SY1tWt9umX
@Accomplish_ai @boundarybench

[Accomplish @Accomplish_ai]

Hi Mark 👋

Congrats on the release, Muse Spark 1.2 is a great model!

You forgot to test it on a real organizational environment, but no worries, we did it with @boundarybench

[Or Hiltch @_orcaman]

Agentic loops are stubborn, which is why agents are effective.

However, in hardened enterprise environments, things start to break.

How do they break, and how can you steer the agent's pivoting behavior the right way?

Some findings from our research, Fable/Mythos included! 👾

[Accomplish @Accomplish_ai]

Accomplish is now on the OSSCAR Index by @commitvc and @supabase as one of the fastest growing open-source projects in Q1 2026.

Open-source doesn’t grow in isolation
Thanks to everyone building, testing, and shaping it.

(we’ve since crossed 10k GitHub stars)

#osscarindex

[Or Hiltch @_orcaman]

Honored to have @Accomplish_ai featured on the official @NVIDIA_AI_PC blog alongside Google’s new Gemma models!

@nvidia has been a valuable partner of hours since we first launched and it’s been a real pleasure working with @asli_sabanci and the rest of the team.

Excited for what’s next!

Link below ⬇️

[Or Hiltch @_orcaman]

Download here: https://t.co/yuWQoSWMQy

What have users been doing with @Accomplish_ai? Here's a sneak peek from our accomplishers community >>

[Or Hiltch @_orcaman]

Today we're launching Accomplish FREE - powered by our new hybrid model router.

Since we launched @Accomplish_ai a few weeks ago, we've been blown away by what users are building with it. Hundreds of thousands of you downloaded the app and took it for a spin, but one thing kept coming up: not everyone wants to bring their own API key just to get started.

So today we're fixing that with free, built-in models - made possible by a massive shift that's happened in just the past few weeks: the rise of fully hosted open-weight models - locally on your @Windows machine with @nvidia @NVIDIAGeForce, on your Mac with @Apple MLX, or on @Accomplish_ai cloud, for FREE.

Our new hybrid routing algorithm dynamically routes between cloud models and models running locally on your machine - optimizing for local execution by automatically detecting your hardware capabilities and each sub-task's complexity: coding, visuals, simple classifications - every LLM call is routed to the best model.

We also brought some of our favorite enterprise features to the free tier: scheduled task dispatch, Google Workspace integration (@googledrive Docs, Sheets, Slides) via the new Google Workspace CLI, native Slack MCP connectivity, and more.

Accomplish FREE is available for macOS, Windows, and Linux. Download, send a task - and boom, it just works with ZERO configuration!

Download link in bio / first comment >>

[Hai Rozencwajg @HaiRozen]

Windows users, the wait is over. You might not have the Mac aesthetic, but thanks to @Accomplish_ai , you finally have Cowork. Welcome to the productivity club!

[Or Hiltch @_orcaman]

We've all been there: as much as we love fixing grandma's PC, it's time for agents to take over.

Today we are launching @Accomplish_ai for @Microsoft @Windows - the first AI Coworker for Windows!

Let AI remove bloat, analyze sheets and take names!

See it/Download below >>

[Or Hiltch @_orcaman]

We've all been there: as much as we love fixing grandma's PC, it's time for agents to take over.

Today we are launching @Accomplish_ai for @Microsoft @Windows - the first AI Coworker for Windows!

Let AI remove bloat, analyze sheets and take names!

See it/Download below >>

[Hai Rozencwajg @HaiRozen]

Wow! This is so cool!!

[Or Hiltch @_orcaman]

Boy, have we got news for you! 🧑‍🍳

The @openwork_ai team is happy to announce:
- Amazon Bedrock integration, contributed by our friends at @awscloud (thanks guys!)
- Native @deepseek_ai integration
- Integration with @openrouter and @LiteLLM!

Here's how what it looks like >>

[Or Hiltch @_orcaman]

היום שחררנו את @openwork_ai, פרוייקט-האקתון נחמד שהצוות הרים ביומיים מאז שאנתרופיק הוציאו את Cowork, שבונה על אותו הקונספט של Computer Use Agent למחשב האישי, אבל עם כמה הבדלים מעניינים:

1. ב- @openwork_ai אתם יכולים להשתמש באיזה מודל שבא לכם, כל עוד הוא נתמך בידי @opencode, החלופה האופן סורסית לקלוד קוד.

2. במקום להכריח את היוזר להתקין chrome extension, אנחנו משתמשים ב dev-browser של @sawyerhood, שהוא מהיר ויעיל פי 4 בערך מתוסף הכרום של אנתרופיק.

3. בטיחות - סוכני CUA חשופים למתקפות prompt injection כשהם משוטטים להם ברשת. בגירסה שלנו הדפדפן לא מחובר לכלום כברירת מחדל, אלא דורש מהיוזר להתחבר לכל שירות שהאייג׳נט צריך לגשת אליו. זה פחות נוח, אבל גם פחות מסוכן.

4. הפרוייקט שלנו חינמי לחלוטין ובקוד פתוח ברשיון MIT, תתפרעו!

לינק ל-GitHub בביו (@openwork_ai). נשמח לשמוע מה אתם חושבים (או שפשוט תשלחו pull request 🙂)

[Hai Rozencwajg @HaiRozen]

Check out what I built with @base44!