Overview

Every skill in Evonic has a loading mode controlled by one field in skill.json: lazy_tools. This setting decides when the skill’s tools and system prompt become available to an agent.

Mode	lazy_tools	Tools available	System prompt added
Non-lazy	false	Immediately at agent start	During system prompt building
Lazy	true	Only after use_skill() is called	Via use_skill()

---

Non-Lazy Skill (lazy_tools: false)

In this mode, the skill’s tools and system prompt are ready the moment the agent starts. It works just like a built-in tool.

Example — Scheduler skill:

{
  "id": "scheduler",
  "lazy_tools": false,
  "tools_file": "tools.json"
}

How it works:

1. When an agent starts, Evonic collects tool definitions from all non-lazy skills.
2. Lazy skills are skipped during this collection.
3. Non-lazy tool definitions go straight into the tools array sent to the LLM.
4. System prompts from non-lazy skills are included in the agent’s system prompt.

Best for: Skills the agent always needs, like file tools or scheduling.

---

Lazy Skill (lazy_tools: true)

The skill stays “asleep” until the agent needs it. Tools and system prompts are only loaded when the agent calls use_skill({id: '...'}).

Example — Kanban skill:

{
  "id": "kanban",
  "lazy_tools": true,
  "tools_file": "tools.json"
}

How it works:

1. Agent calls use_skill({id: 'kanban'}).
2. The tool checks the skill is enabled, allowed, and reads the SYSTEM.md file.
3. If lazy_tools is true, it also loads the skill’s tool definitions.
4. Tools are injected into the LLM context mid-turn — the agent can use them right away.
5. The skill’s system prompt is stored and re-injected every loop iteration so it doesn’t get lost in long conversations.

To remove a lazy skill, call unload_skill({id: '...'}). This removes its tools and system prompt from the context.

Best for: Optional or rarely-used skills that don’t need to take up LLM context space all the time.

---

Browsing Available Skills

Agents can see what skills are available through a section in the system prompt. The ## Skills section lists all skills (both lazy and non-lazy) with short descriptions:

## Skills
You have these skills that can be loaded using `use_skill` tool:
- `kanban` - Allows agents to interact with the Kanban board...
- `subagent` - Spawn ad-hoc sub-agents...

The difference:

• Non-lazy skill — tools are ready immediately.
• Lazy skill — tools need use_skill() to load.

---

How use_skill and unload_skill Work

use_skill and unload_skill are built-in tools available to every agent. They don’t need to be loaded from a skill — they’re always there.

• use_skill — loads the skill’s SYSTEM.md content and, for lazy skills, injects tool definitions.
• unload_skill — removes the skill’s tools and system prompt from the agent’s context.

Skill state (what’s loaded) is tracked per-session. If an agent loads a skill in one turn, it stays loaded in the next turn — until unload_skill is called.