Agents

An agent is a configurable AI assistant. It combines a language model with instructions, optional tools, memory, and knowledge — packaged as a reusable unit that the Alquimia Runtime can execute.

What makes up an agent

Component	What it does	Required
Language model	The LLM that generates responses	Yes
System prompt	Identity, personality, rules, and purpose	Yes (auto-generated or custom)
MCP tools	External services the agent can call	No
Agent-to-Agent	Other agents the agent can delegate to	No
Memory	Conversation history (short-term) or persistent context (long-term)	No
Knowledge base	Document collection for RAG-based answers	No
Integration channels	WhatsApp, Email — external communication interfaces	No
Pre-processor	Sentinels and empathy rules that run before the agent	No

How agents run

When a user sends a message:

User message
    ↓
Pre-processor (sentinels, empathy rules) — optional
    ↓
Memory retrieval — inject relevant history
    ↓
Knowledge retrieval — inject relevant document passages
    ↓
LLM call — prompt + context + tool schemas
    ↓
Tool execution (if needed) — ReAct loop
    ↓
Response returned to user

The Runtime handles all of this automatically based on the agent’s configuration. You don’t write any of this logic — you configure it in Studio.

Evaluation strategies

The evaluation strategy controls how the agent thinks. This is one of the most important settings for an agent.

Strategy	Description	Best for
`off`	Single LLM call, no reasoning loop	Simple Q&A, content generation
`one-shoot`	One call with tool schemas injected	Structured output, simple tool use
`tools`	LLM calls tools in a single round	Deterministic, tool-first workflows
`react`	ReAct loop: reason → act → observe → repeat	Multi-step tool use, complex tasks
`long-task`	Extended planning for complex, multi-step tasks	Research agents, autonomous workflows

When you add MCP servers or other agents to your agent, Studio automatically switches the evaluation strategy to react. This ensures the agent can actually use its tools. You can override this manually in Dev Mode.

The ReAct loop explained

react is the most powerful strategy. The agent follows this cycle:

Reason — “I need to look up the user’s account to answer this”
Act — calls the get_account tool from an MCP server
Observe — receives the tool result
Reason again — “Now I have the account info, I can answer”
Respond — generates the final answer

This continues until the agent has enough information to respond or hits its reasoning limit.

Welcome

Getting Started

Platform

Core Concepts

Ecosystem

Reference

What makes up an agent

How agents run

Evaluation strategies

The ReAct loop explained

Next steps

Creating an Agent

Dev Mode

​What makes up an agent

​How agents run

​Evaluation strategies

​The ReAct loop explained

​Next steps

Creating an Agent

Dev Mode

What makes up an agent

How agents run

Evaluation strategies

The ReAct loop explained

Next steps