Video yükleniyor...
Video Yüklenemedi
🚀 PilotDeck 2.0 is here! Meet the open-source AI Agent Harness built to get complex work done. Developed by Tsinghua THUNLP, ModelBest, OpenBMB , and AI9Stars, PilotDeck is a model-agnostic Agent operating system that turns natural language instructions into completed tasks — with transparent data flows and the freedom... show more
549,573 görüntüleme • 1 gün önce •via X (Twitter)
32 Yorum
@modelbest_ai another day, another reason to love open source
@modelbest_ai an open source harness for agents is the real model release
@modelbest_ai the fact that you can actually see what the agent is doing is a huge plus.
@modelbest_ai OpenBMB, the harness that shipped is the one where the instruction already became the completed thing.
@modelbest_ai Exactly! 🙌 Less talking about tasks, more actually getting them done. That's what we're building toward with PilotDeck! 🚀
@modelbest_ai the real test is whether i can trust it with my spreadsheets
@modelbest_ai this would save me so much back and forth if it works as advertised
@modelbest_ai open source agents are getting seriously interesting
@modelbest_ai Less chatting, more doing. that's the whole point
@modelbest_ai ngl, i'd love to see someone push this with a really messy real-world task
@modelbest_ai Can it ask for approval before overwriting an existing file?
@modelbest_ai The more capable agents become, the more important it is to make their work inspectable: automation earns trust when people can see what happened and intervene.
@modelbest_ai model agnostic is the part people will actually care about, nobody wants their harness married to one lab
@modelbest_ai the first real-world task: update three slides, keep the speaker notes, don't break the theme
@modelbest_ai thanks, 79.1 on an A3B model
@modelbest_ai Hold up, how does it recover if something breaks halfway through?
@modelbest_ai @goog
@modelbest_ai Would love to see cost per completed task alongside the benchmark score.
@modelbest_ai The shift from drafts to verified office files is the real unlock: people can hand agents a full task and still check the result.
@modelbest_ai finally something for people who don't spend their entire day coding
@modelbest_ai I need to throw my messiest work project at this
@modelbest_ai office files are where an agent's mistakes hide. a wrong formula looks fine until someone checks the total. verifying helps, but being able to reverse the edit is the part i'd test first.
@modelbest_ai 79.1 with Qwen?? that's pretty wild
@modelbest_ai nice. now make it survive my boss changing the requirements 12 times 😂
@modelbest_ai ok but can it handle a spreadsheet with 15 tabs and broken formulas 💀
@modelbest_ai can we make model agnostic the default for AI tools already
@modelbest_ai Model-agnostic is the right call, but the hard part is usually the execution layer: browser actions that break on a layout change or a slow render. Curious how PilotDeck handles retries and state when a step fails mid-task.
@modelbest_ai can it fix my 47-slide powerpoint at 2am tho
@modelbest_ai Beating Claude Code on Workbuddy Bench with a 35B-A3B model is a big claim. How much of that is the harness vs the model? Did you also run Claude Code with Qwen3.6, so it's apples to apples?
@modelbest_ai PilotDeck 2.0 is a huge step toward truly autonomous AI agents that don’t just generate content but actually get complex work done! 🚀
@modelbest_ai An undo button for a whole agent run would be very welcome here.
@modelbest_ai wait, it can actually edit the files too??
