Google has turned the Gemini interface into a real “command center” by introducing subagents into it, xrust reported. Now, instead of one neural network, a whole team of highly specialized AI assistants can work on a task.
Previously, working with large projects in AI was like trying to keep thousands of details in your head at once. If you load too much data into the system, a “blurring effect” occurs: the neural network begins to get confused, forgets important instructions and produces contradictory results.
Engineers call this “context decomposition.” When there is too much information, the extra “noise” simply drowns out important data. To solve this problem, Google has implemented a system of subagents that can be called with a simple command via the @ symbol.
How does this work in the example of a construction site?
To develop a house project:
- The “main neural network” receives a common task and coordinates the process.
- The assistant is engaged in excavation work and foundation design.
- Other assistants work in parallel with utility networks, heating and ventilation.
Each “subagent” works in its own isolated zone. He sees only those drawings and data that he needs.
What are the benefits?
When a developer needs to rewrite the database schema, he calls @db-agent. This assistant does its part of the job without “cluttering” the memory of the main model with unnecessary details. As a result, the main Gemini session remains fast and focused on the main goal, and the result is much more accurate.
Based on materials from https://www.developer-tech
Xrust Google has updated the Gemini neural network: now AI has narrow-profile «helpers»
