Intelligent agents are the ultimate goal of AI. These are systems that can perceive their environment, plan, and act to achieve a goal. With the rise of foundation models, this vision is finally becoming reality.
What is an AI Agent?
An agent is anything that can perceive its environment and act upon it. In the context of AI, an agent uses a language model as its brain to process the task, plan a sequence of actions, and determine if the goal has been accomplished.
The Role of Tools
A model by itself is limited; it can only generate text. Tools extend its capabilities. Tools allow an agent to perceive the environment (like a web search or SQL query) and act upon it (like sending an email or updating a database).
Examples of tools include:
- Knowledge Augmentation: Text retrievers, SQL executors, and web browsers.
- Capability Extension: Calculators, code interpreters, and image generators.
- Action: Email APIs, bank transfer APIs, and Slack posting tools.
The key challenge is giving an agent a rich inventory of tools without making it too complex to use them effectively.
Planning and Reflection
Complex tasks require planning. The agent must decompose a task into a sequence of manageable actions. However, LLMs are notoriously poor at planning. They often hallucinate steps.
A better approach is to decouple planning from execution. The agent generates a plan, which is validated by heuristics or an AI judge, and only then executed. Furthermore, the best agents utilize reflection. After each action, the agent evaluates the outcome. If the result is unsatisfactory, the agent reflects on the mistake and proposes a new plan.
Evaluation
Evaluating agents is difficult. There is no single standard. You must measure:
- Success Rate: Does it accomplish the task?
- Cost: How many tokens are consumed?
- Speed: How long does it take?
Human evaluation is still required to catch errors like hallucinations.