The Next Breakthrough In AI Agents Is Here

Garry Tan breaks down Manus, a multi-agent AI system scoring 86.5% on GAIA, and argues why ‘wrapper’ startups can still win.

• Manus scored 86.5% on the GAIA benchmark, beating OpenAI Deep Research (74%) and approaching average human performance (92%).
• Manus costs ~$2 per task, significantly undercutting integrated competitors like OpenAI’s Deep Research.
• Architecture uses a planner agent that decomposes tasks, 29 integrated tools, and a chain-of-thought injection technique to maintain stability across long reasoning chains.
• Manus is built on Claude 3.7 Sonnet and integrates YC-backed Browser Use and E2B’s cloud sandbox.
• Co-founder Yichao Peak G explicitly chose to work orthogonal to model development, treating new model releases as an advantage rather than a threat.
• Garry Tan argues successful wrappers win through proprietary evals, fine-tuning, sticky UX, and data integrations competitors can’t replicate—not by avoiding the wrapper label.
• Key wrapper risks: vulnerable to API pricing changes, provider policy shifts, and fast-follower competitors copying UX refinements.

2025-04-08 · Watch on YouTube