Agent types

Plain LLM Agents

Plain LLM agents are simple agents that use a large language model (LLM) directly to choose next actions or generate task steps without extra layers like complex planners, learned policies, or specialized orchestration frameworks.

Key characteristics:

• Single-step decisioning: the LLM is prompted to decide the next action each turn (e.g., call a tool, ask a question, produce text).
• Minimal state management: little or no explicit memory, belief model, or long-term planning beyond what’s kept in the prompt/history.
• No learned controller: decisions rely on prompt engineering and the LLM’s reasoning, not on a separate trained policy network.
• Tool-driven behavior: often constrained to a fixed set of tools or API calls the LLM can invoke via structured outputs.
• Reactive and iterative: acts, observes results, and prompts the LLM again—adapting only through updated context.

When to use:

• Prototyping agents quickly.
• Tasks where short-horizon, conversational reasoning suffices.
• Systems prioritizing simplicity and interpretability.

Limitations:

• Poor scalability for long, complex plans.
• Fragile to prompt drift and verbose histories.
• Limited ability to optimize across multiple steps or maintain consistent long-term strategies.