AI Agents

Configurable AI agents that use a ReAct (Reasoning + Acting) loop to complete multi-step tasks using tools.

Use case: An agent tasked with "What's 15% of the current Bitcoin price?" will search the web for the price, then use the calculator tool to compute the result — all automatically.

How It Works

1. Agents are defined with a name, provider, model, system prompt, tool list, temperature, and iteration limits.
2. When run, the agent sends the task to the LLM with tool definitions.
3. If the LLM calls a tool, the tool is executed and the result is appended to the conversation.
4. The loop continues until the LLM returns a final answer or maxIterations is reached.
5. Steps (thinking, tool call, tool result, final answer) are returned for inspection.

Available Tools

Tool	Description
get_datetime	Returns current ISO 8601 datetime, UTC string, and Unix timestamp
calculate	Safely evaluates math expressions (+, -, *, /, %, parentheses)
web_search	DuckDuckGo instant-answer search (no API key required)
http_get	Makes HTTP GET requests to public URLs; returns up to 4000 chars of body
database_query	Runs read-only SELECT queries on the Rotifex SQLite database

ReAct Loop

User Input
    |
    v
LLM (receives task + tool definitions)
    |
    +-- Tool call requested?
    |         |
    |         +-- Execute tool
    |         +-- Append result to conversation
    |         +-- Loop back to LLM
    |
    +-- Final answer -> return output + all steps to caller

Agent Step Types

Type	Description
thinking	LLM reasoning text before invoking a tool
tool_call	Name and arguments of the tool being called
tool_result	Raw output from the tool execution
final_answer	The LLM's conclusive response