Angelina YangFounder and Host, TwoSetAI · AI builder for visibility and discoverability

How AI Handles 10 Million Tokens Using Only 150 Agents

4 Aug 2026 · 1 min

Watch on YouTube

0:45

In this brief interview excerpt, Speaker A asks how the system can manage up to 150 agents per task when faced with extremely large contexts, potentially exceeding 10 million tokens. Speaker B explains that a single model cannot hold such massive context (over 1 million tokens). Instead, the approach uses a divide‑and‑conquer strategy: the overall problem is broken into smaller subtasks, each handled by a separate search/agent that only needs a much shorter, task‑specific input. The example given is locating a 1980 calculus textbook, which requires only a concise query rather than the full multi‑million‑token dataset. This illustrates how the architecture leverages many lightweight agents to solve problems that would otherwise exceed a single model’s context window.

I read in your blog that you have up to 150 agents for per task. How do you handle the large amount of context?
A single agent or a single large model cannot hold a very, very long context, say more than 1 million. If you basically do divide and conquer, then you can decompose a very hard problem into subtasks. The original problem, it may take more than 10 million tokens to solve. What you actually need as input to search agent, it might be much simpler. For example, find a math book in 1980 about calculus. For this specific problem, you don't need very, very long content.