7 Practical Design Patterns for Agentic Systems
In the generative AI era, AI agents represent systems where LLMs autonomously perform complex tasks from start to finish. Agentic AI systems however, represent a more holistic approach to LLM autonomy by encompassing systems that lie somewhere between simply prompting an LLM and fully autonomous AI agents. These systems range from implementations where LLMs make limited decisions within a structured workflow to those where LLMs independently execute tasks with minimal human intervention. The degree of "agenticity" is defined by how much decision-making authority is delegated to the LLM.
Some common design patterns for agentic systems are as follows:
- Controlled flows
- LLM as a router
- Parallelization
- Reflect and critique
- Human in the loop
- Agents
- Multi-agent
The “right” design pattern depends on the core requirements of your application. Some parameters to consider are reliability of responses, structuredness of the workflow, complexity of the tasks to be performed, and consequences of bad outcomes.
Start with the simplest architecture that could work, carefully evaluate its performance, and add additional components only if there is clear evidence that they are needed.