Nishant Soni on the evolution of AI agents (or agentic AI) toward “agencies”:
What I’m witnessing is the birth of what I believe should be called Agencies - systems that tackle individual tasks by dynamically orchestrating different types of intelligence, each optimized for specific subtasks, all working toward completing a single overarching objective. An Agency is fundamentally different from an Agent. While an Agent is a single intelligence (an LLM) enhanced with tool-calling capabilities working on a task, an Agency is a coordination system that can access and deploy multiple specialized intelligences (LLMs) to complete different parts of the same task.
Think of an agency as a “boss AI” (a term I learned from Scot Wingo, founder and CEO of ReFiBuy.ai – he brought this up on our recent podcast conversation), consisting of three parts:
1. Task Context Management: The Agency maintains unified context about the specific task at hand - requirements, constraints, progress, and accumulated decisions. This ensures continuity as different intelligences contribute to different subtasks.
2. Intelligence Allocation System: Rather than using one model for everything, the Agency has access to multiple specialized intelligences and dynamically selects the most appropriate one for each subtask within the larger task.
3. Orchestration Logic: A coordination system that breaks down the main task into subtasks, determines which intelligence to use for each part, and ensures all contributions integrate coherently toward task completion.
In summary: “Agencies are not multiple Agents collaborating on a project. They are single unified systems that can access multiple types of intelligence to complete individual tasks more effectively. […] We’re moving beyond asking ‘What’s the best model for this task?’ to ‘What’s the best combination of intelligences for different aspects of this task?’”
In many ways, this is the evolution of “mixture of experts,” where a single AI (LLM) has access to multiple, specially trained, typically smaller models and routes requests to the most capable model for a specific task (e.g., a model which is optimized for coding tasks).
