Agents and Skills¶

Understanding the relationship between Agents and Skills is the key to mastering OLAV.

This is a conceptual overview. For related Feature Claims, see Using Agents, Build a Skill, and Manage Workspace.

Agent: An AI Assistant Focused on a Specific Task¶

Think of Agents as different roles on a team. Each Agent has a clear responsibility and dedicated tools, and excels at handling a specific type of problem.

An Agent consists of three parts:

Component	Description	Corresponding File
System Prompt	Defines the Agent's role, behavioral guidelines, and output style	`system_prompt_file` in `AGENT.md`
Toolset	Functions the Agent can call (query databases, invoke APIs, execute code...)	Loaded via SubAgents and Skills
Route Keywords	When a user's question matches these keywords, OLAV automatically selects this Agent	`route_keywords` in `MANIFEST.yaml`

AGENT.md Example¶

---
name: quick-orchestrator
description: "Quick Query Agent — fast SQL/CLI queries"
system_prompt_file: prompts/system.md
subagents: []
static_context:
  - path: ./references/SCHEMA_REFERENCE.md
---

## Overview
Quick Agent is the default entry point for OLAV.
It skips complex multi-step reasoning loops in favor of immediate, single-step responses.

The static_context field specifies reference materials (such as database schemas) that the Agent automatically loads on each run, helping the LLM understand the problem more accurately.

Skill: A Toolkit That Gives Agents New Capabilities¶

A Skill is a collection of Python @tool functions that can be independently developed, installed, and shared. Skills fall into three categories by origin:

Type	Description	Example
Built-in	Shipped with OLAV	`core` (code execution), `config` (system management)
Generated	Auto-generated from OpenAPI by the Creator Agent	Zabbix tools, NetBox tools
Custom	Written by you, installed via `olav skill install`	Ticket queries, automation scripts

MANIFEST.yaml Example¶

kind: Agent
name: quick
version: "0.11.0"
description: "Quick Agent — fast SQL queries, CLI execution, knowledge base search"
route_keywords:
  - quick query
  - show me
  - list devices
  - what is
  - how many
tools_dir: tools/
requires:
  - olav-platform>=0.11

route_keywords determines when the semantic router assigns a question to this Agent. tools_dir specifies the directory containing the tool functions.

How Agents Load Tools¶

Agents acquire tools in two ways:

Static Binding (Declared in AGENT.md)¶

subagents:
  - path: ./creator/SKILL.md
  - path: ./knowledge/SKILL.md

Suitable for built-in, stable tool bindings. The Agent loads these sub-skills every time it starts.

Dynamic Discovery (Via `olav skill install`)¶

olav skill install ./my-tools/

After installation, tools are written to .olav/workspace/<name>/. When OLAV starts an Agent, it automatically scans the workspace, discovers and injects all installed skill tools. No manual modification of AGENT.md is needed.

Tool Call Loop¶

When you ask a question, the Agent's internal processing flow is as follows:

Your question
    |
Agent sees all available tools (each with a name and description)
    |
LLM decides which tool to call and passes in arguments
    |
Tool executes and returns results
    |
LLM analyzes results and decides whether more tools are needed
    |
Once all information is sufficient, synthesizes a final answer
    |
Each tool call is automatically written to the audit log

Built-in Agents Overview¶

Agent	Role	Strengths
`quick` (default)	Quick query assistant	Single-step queries, one question one answer
`config`	System administrator	Service registration, skill generation, workspace health checks
`core`	Full-stack engineer	Code execution, SQL, Shell, web search

olav list   # View all available Agents

When Should You Create What?¶

Need	Solution
Add new capabilities to an existing Agent	Install a new Skill
Need a completely different AI assistant role	Create a new Agent (new AGENT.md)
Add a few tools to an existing Skill	`olav skill install --merge-into`

In most scenarios, you only need to install a new Skill. You only need to create a new Agent when you require a different system prompt or different behavioral patterns.