Intelligent agents are considered by many to be the ultimate goal of AI. The unprecedented capabilities of foundation models have opened the door to agentic applications that were previously unimaginable. These new capabilities make it finally possible to develop autonomous agents to act as our assistants, coworkers, and coaches. They can help us create websites, gather data, plan trips, and more. The potential economic value is enormous.
This guide covers the foundational concepts of agents, focusing on two key determinants of their capabilities: tools and planning. It also explores the new modes of failure that agents introduce and discusses how to evaluate them.
What is an Agent?
An agent is anything that can perceive its environment and act upon it. It is characterized by the environment it operates in and the actions it can perform. In an AI agent, the AI model is the brain that processes the task, plans a sequence of actions, and determines if the task has been accomplished.
Tools: The Agent's Inventory
An agent's capabilities depend heavily on its tool inventory. Tools help an agent both perceive its environment and act upon it. Without external tools, an agent's capabilities are limited.
Categories of Tools
- Knowledge Augmentation: Tools like text retrievers, SQL executors, and Slack retrieval to provide context.
- Capability Extension: Tools to address limitations, like a calculator for math, or a code interpreter.
- Read and Write Actions: Tools can retrieve data (read) or modify it (write), like an email API that sends messages.
Tool use can significantly boost performance. The Chameleon system, for example, uses 13 tools to outperform GPT-4 alone on science and math benchmarks.
Planning: The Roadmap
Complex tasks require planning. The output is a roadmap outlining the steps needed. Effective planning involves understanding the task, considering options, and choosing the most promising one.
Decoupling Planning and Execution
To avoid wasting resources on a bad plan, decouple planning from execution. The agent first generates a plan. The plan is then validated using heuristics or AI judges. Only after validation is it executed. This process often involves plan generation, validation, and execution, creating a multi-agent system.
Control Flows
Plans can be sequential, parallel, conditional (if statements), or iterative (for loops). The choice of control flow impacts latency and cost. Parallel execution, for example, can significantly reduce user-perceived latency.
Reflection: Learning from Mistakes
Reflection is a crucial, though not mandatory, component. It involves evaluating outcomes and correcting mistakes. Frameworks like ReAct interleave reasoning and action. Reflexion adds a dedicated self-reflection module to analyze what went wrong. While this improves performance, it increases latency and cost.
Evaluating Agents
Evaluation is about detecting failures. Agents have unique failure modes in planning, tool execution, and efficiency. Common planning failures include hallucinating invalid actions, using wrong parameters, and failing to achieve the goal. Tool failures occur when a tool returns wrong outputs. Efficiency metrics, like average action time and number of steps to success, are also critical.
At its core, the concept of an agent is simple. An AI model leverages tools and feedback to plan tasks. Access to tools makes a model vastly more capable, making the agentic pattern inevitable.