Click on each principle to expand and learn more.
To get the best results, you need to understand how the AI agent "thinks." This means having a clear mental model of its capabilities, its limitations, and the way it processes information. Crafting prompts that align with this model leads to more accurate and relevant answers.
For complex research, a single AI can't do everything well. An orchestration agent acts as a project manager, breaking down the main query into sub-tasks and delegating each to the best specialized agent or tool. The accuracy of this orchestrator is essential for a correct final output.
The system's effort should match the difficulty of the question. A well-designed system scales its effort automatically, ensuring efficiency for simple tasks while dedicating the necessary resources to solve complex problems.
A research agent's effectiveness hinges on its ability to choose the right tool for the right job (e.g., web search, calculator, code interpreter). Accuracy in tool selection is vital; the agent must correctly map the task to the most appropriate tool.
Effective research starts broad before getting specific. The AI should first perform a general search to explore the landscape of a topic. After mapping the area, the agent can then perform more targeted, "drill-down" searches for specific facts or data points.
Users and system designers should guide the AI's thinking process, often by structuring the prompt as a chain of thought. This structured guidance helps the agent stay on track, build on previous steps logically, and avoid errors.
AI systems should learn from their interactions to get better over time. This is known as a Data Flywheel. Every task generates valuable data that is used to fine-tune the models, allowing the agent to improve itself with each use.
When a task involves multiple independent steps, they should be executed simultaneously. Parallel tool calling allows the agent to run multiple searches or operations at once, which dramatically reduces waiting time and improves performance.