AI Agents: Overview and Core Concepts

AI Agents in Flametree are configurable logic units powered by large language models (LLMs). They process user inputs, execute actions, and manage context across workflow states using tools, skills, and observations.

Core Components

Each agent combines reasoning with structured execution through three core elements:

Tools

External integrations that perform concrete actions — such as fetching data, sending messages, or validating input.

Skills

Reusable logic blocks triggered by workflow events or state transitions. They define how the agent behaves under specific conditions.

Observations

Data returned from tools or previous actions. Observations are stored in context and influence the agent’s subsequent decisions.

Together, these components form a continuous loop:

Input → Reasoning → Action → Observation → Next Decision

How Agents Operate

Agents run inside workflows that define states, transitions, and logic flow. During execution, an agent:

1. Receives Input — from a user message or external event.
2. Evaluates Context — checks current state and previous observations.
3. Executes Tools or Skills — performs configured actions.
4. Processes Observations — interprets returned data.
5. Responds or Transitions — sends a message or moves to the next workflow state.

This architecture allows AI Agents to combine flexible language reasoning with deterministic state management, ensuring predictable and traceable behavior.

Example: Appointment Booking Agent

A booking agent might use a tool CheckAvailableDates to retrieve available time slots:

• The agent calls CheckAvailableDates and stores results as observations.
• It generates a message like:
  “Available dates are May 3rd, 4th, and 5th. Which one would you prefer?”
• After receiving a reply, it updates state and proceeds to confirmation.