The term “agent” is used loosely across the AI industry. Some definitions stretch to include any system that calls an LLM; others reserve the word for fully autonomous programs that operate independently for extended periods. This page offers a working definition grounded in Anthropic’s research on building effective agents.
A Working Definition
An AI agent is a system where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks rather than following a fixed, predetermined code path. The key word is dynamically — the model decides at each step what to do next based on the results of its previous actions.
This stands in contrast to systems where a human-written program orchestrates every call to the LLM in a predefined sequence. Both approaches are valid. The distinction matters because it determines where decision-making authority lives: in your code or in the model.
The Spectrum of Complexity
It helps to think of LLM-based systems as sitting on a spectrum, ordered by how much autonomy the model has:
-
Simple LLM call. A single prompt goes in, a single response comes out. No tools, no loops. This covers a surprising number of real-world use cases.
-
Augmented LLM. The model is enhanced with access to retrieval systems, external tools, or memory. It can pull in context or take actions, but each interaction is still relatively contained.
-
Workflows. Multiple LLM calls are chained together through predefined code paths. The orchestration logic is written by a developer — the LLM executes specific roles at specific points.
-
Autonomous agents. The LLM controls its own loop. It decides which tools to call, in what order, and when to stop. Human involvement is minimal once the agent is launched.
Most production systems today sit somewhere in the first three categories. Fully autonomous agents are powerful but harder to debug, predict, and control.
The Augmented LLM: The Basic Building Block
Regardless of where a system falls on the spectrum, the augmented LLM is the fundamental unit of construction. An augmented LLM combines a base language model with one or more of the following capabilities:
- Retrieval. The ability to pull in relevant information from external sources — documents, databases, APIs — so the model can reason over knowledge it was not trained on.
- Tools. The ability to take actions in the world: run code, query a database, call an API, manipulate files.
- Memory. The ability to persist information across interactions, whether through conversation history, vector stores, or structured state.
Getting these augmentation layers right is essential. A well-tuned retrieval system or a carefully designed tool interface often delivers more practical value than a complex multi-agent architecture built on shaky foundations.
The Guiding Principle
Anthropic’s research emphasizes a single, recurring theme: keep your design as simple as possible, and add complexity only when simpler solutions demonstrably fail. An optimized single LLM call with good retrieval will outperform a poorly designed agent on most tasks. Complexity should be earned through measured performance gains, not assumed in advance.
The pages that follow unpack how to think about the architectural choices between workflows and agents, and when each is the right tool for the job.