Pinned
After four overhauls and millions of real-world sessions, here are the lessons we learned about context engineering for AI agents:
Yichao 'Peak' Ji
66 posts
Yichao 'Peak' Ji
@peakji
Joined February 2010
- Replying to @jianxliao and @browser_useHi! I'm Peak from Manus AI. Actually, it's not that complicated - the sandbox is directly accessible to each user (see screenshot for method). Specifically: * Each session has its own sandbox, completely isolated from other sessions. Users can enter the sandbox directly through
- Actually, Manus doesn't use MCP. We were more inspired by my friend @xingyaow_ 's work: openreview.net/forum?id=jJ9Bo…. While we haven't fully adopted CodeAct, this work provided 3 crucial insights: * Coding is not the ultimate goal, but rather a universal approach for solving general
- Feeling so much love from the community today! Manus launched with rumors flying, and we panicked. But as more people tried it (crashes and all), the truth spoke for itself. People I quietly admired reached out to help—it felt unreal. Grateful beyond words for this support. ❤️
- Important Notice Regarding Manus's X Account Suspension Our official X account (@manusai) was unexpectedly suspended yesterday. We're actively working with X's support team to resolve this matter. Preliminary observations suggest the suspension might relate to third-party
00:00 - Good news: We've nearly doubled @manusai 's context limit & beefed up our sandbox hardware. Failure rate down dramatically! 📉 Bad news: With lower failure rate, now sessions run longer, system's sweating bullets, and failures are climbing back up... Had to throttle things
- The Jevons Paradox in AI is absolutely real. Over the past two years, we've been building chatbots and have witnessed the inference costs of models with the same level of intelligence decrease by 50 to 100x. Last year, when we believed that the efficiency was sufficient to
- Replying to @TenzinTheCyber @jianxliao and @browser_useWe use Claude and different Qwen-finetunes. Back when we started building Manus, we only got Claude 3.5 Sonnet v1 (not long-CoT, aka reasoning tokens), so we need a lot of auxiliary models. Now Claude 3.7 looks really promising, we are testing internally, will post updates!
- Replying to @heyBarseeCo-founder of Manus AI here! Thanks for your interest in Manus @heyBarsee, but this video is definitely NOT Manus🤣. Our official account is @manusai , and Manus is a web-based AI agent, like OpenAI Operator/DeepResearch, not a VNC for smartphones.
- Replying to @toprak_dikici @jianxliao and @browser_useWe have a lot of fine-tuned models in our company. Some Manus' planning technique originated from this model I've open-sourced last October:
- Thanks, Rowan! And we just had our Rowan moment.I think China's second DeepSeek moment is here. This AI agent called 'Manus' is going crazy viral in China right now. Probably only a matter of time until it hits the US. It's like Deep Research + Operator + Claude Computer combined, and it's REALLY good.
00:00 - We are back! This is the ONLY official account of Manus ⬇️ Just to be clear: - Manus has never been involved with cryptocurrency projects, token issuance, or blockchain initiatives - Any similarly named cryptocurrency ventures claiming association with us are fraudulentWould you like ask Manus to complete a task for you? Reply to this tweet with a task you'd like Manus to handle. Our team will run your request with Manus and share the results with you. Whether it's research, content creation, data analysis, or something creative - let's see
- Today a16z released the latest AI Top 50 list. As a small startup, it's truly an honor that we made it with not just one, but two products. The company is charging toward the $100M milestone, and I'm about to enter my 15th year in entrepreneurship. In the beginning, I thought
