Nov 11, 2024
Our classification combines several classical approaches to agent systems with modern developments in AI. Historically, agent classification emerged from robotics and AI research in the 1990s-2000s, introducing concepts like agent types, architectural patterns, and autonomy levels. With the advent of LLMs, these concepts have evolved to embrace new capabilities.
Autonomy levels
Level 0: Manual Control: No autonomy. Human controls/programs all actions
Level 1: Decision Support: Basic autonomy. The system suggests — human decides
Level 2: Consensus: Moderate autonomy. The system acts independently but requires approval for critical decisions
Level 3: Supervised Autonomy: High autonomy. The system acts independently with human oversight
Level 4: Full Autonomy: Full autonomy. The system can make all decisions on its own. Humans only set high-level goals and constraints.
Agent types
Reactive: Simple stimulus-response behavior without an internal world model
Deliberative: Uses internal world model for planning and decision-making
Hybrid: Combines both reactive and deliberative approaches for optimal performance
Let's start with basic definitions
Automation: Execute predefined, rule-based tasks automatically. This includes both traditional software (scripts, programs) and ML integrations. These are typically reflex agents that perform one certain task well and consistently.
Uses reactive architecture and operates at Level 0-1 autonomy, where all significant decisions require human input or follow strict rules.
Examples: script for backup or CI/CD pipeline, which simply executes given commands and behavior fully defined by humans.
AI Workflow: Predefined paths that execute tasks sequentially. More complex rule-based agents can use current perception and draw on memory, enabling them to receive and store new information. It can include LLM integration for specific steps but maintains predetermined execution paths. However, they usually rely on machine learning models cause they are much faster and cheaper than LLMs. If you have data and a person to train them.
Shows Level 1 autonomy with basic decision support capabilities within predefined boundaries.
Examples: Antivirus, suggesting actions with a suspicious file based on ML model predictions.
AI Agent: Dynamic, self-directed systems designed to perform non-deterministic tasks autonomously. Usually LLM-based, as it requires complex context understanding and decision-making. Can actively gather or use new information beyond training data with RAG. Could interact with external tools and services for that purpose. Usually, there is no feedback loop for results; do something and rely on humans to evaluate results. Operating at Level 3-4 autonomy, capable of independent decision-making with human oversight. Examples: ChatGPT, which can answer questions, generate text, and interact with external APIs, but don`t evaluate results it produces.
Agentic Systems: Solutions that utilize one or more agents with tailored backend and UI with the main idea of providing agents with an interactive environment where they could perform tasks and evaluate results. Implements sophisticated layered and hybrid architectures, allowing multiple agents to cooperate. Each agent can operate at different autonomy levels (2-4) depending on their role and task criticality. Such a system utilizes the idea of creating complex behavior by combining simple agents. Could operate up to Level 4 autonomy. Examples: Ardor, Cursor, Replit. These have interactive environments for agents with access to different tools and data from linter, shell, git, etc.
Let's talk more about Agentic Systems
The main idea of agentic systems is autonomous or partially autonomous work with tasks that are almost impossible to determine.
They have an interactive environment with tools they could utilize depending on the situation. Their environment is tailored for AI agents. In some way, they could be called agents first, as a major part of their functionality exists to provide agents with the necessary information and level of interaction with real-world data.
In this case, humans only set high-level tasks and receive the results.
Let's look at examples
AI Automation
A system sends emails with basic personalization based on user form data. Data, interface, templates are fixed. The only variable is user data, which is used by a generative model to create more personalized messages compared to simple template with user's name.
AI Workflow
A system sends emails considering additional data that works as logical gates - if-then-else. Models within the workflow help to choose user cohorts, templates, content recommendations, etc. Usually, classical ML models or LLMs are used to test approaches before training tailored ML models. But they don't perform orchestration and don't call tools.
AI Agent
The system can communicate with users via chat or email. It can request additional data by calling necessary tools (check user balance or order history). Such system can perform non-strictly determined tasks within available tools and data. Could store history of conversation to be more personalized.
Agentic System
A CRM system with multiple agents that can interact with users (internal and external) in various formats. Agents have access to various tools and data that can be used to solve tasks. They help set proper configurations for CRM itself and interact with real-world data. They can also make changes to the system structure - whether creating database migrations or adding new tools to solve tasks. That is, besides the agents themselves, they are in an environment that allows them to interact with real data, users, and tools and add necessary functionality.
Interactive environment for agents:
Discover Tools: Dynamically find and learn how to use new tools
Access Data: Query and manipulate both structured and unstructured data
Modify Environment: Add new tools, update configurations, create workflows
Monitor Results: Track outcomes and adjust strategies
Communicate: Exchange information with other agents, systems, and humans
For example, Ardor is an environment and platform for creating agentic solutions. It has a Copilot (agentic system) that helps to create AI Agents. Since Copilot has vertical integration with Ardor, it can use all the tools and data available in Ardor, including having access to its own architecture (not a model, but a system with all tools and data). This allows it to supplement it with new tools and functions, depending on the task that needs to be solved.