Skills

A skill is a reusable prompt package that injects specific domain knowledge, procedures, and scripts into an agent. Skills are Google ADK–compatible: load local skills through ADK’s SkillToolset, and connect cloud skill spaces through VeADK’s VeSkillRegistry.

Local skills

Load a local skill directory with ADK’s load_skill_from_dir and pass it as a standard toolset to the agent:

from google.adk.skills import load_skill_from_dir
from google.adk.tools.skill_toolset import SkillToolset
from veadk import Agent

skill = load_skill_from_dir("/abs/path/to/skills/kb-skill")
agent = Agent(tools=[SkillToolset(skills=[skill])])

Skill loading, prompt injection, and tool exposure are all handled by ADK; the tools visible to the model include list_skills, load_skill, load_skill_resource, and run_skill_script. If a skill runs scripts from scripts/, configure a suitable code_executor on the SkillToolset or the agent when constructing SkillToolset directly through the ADK API.

When skills are loaded through Harness, CLI-generated agent code, or the AgentKit session capability service, VeADK automatically configures a local code executor for SkillToolset, so skill scripts run without additional configuration.

Cloud skill spaces

Connect a cloud skill space through VeSkillRegistry, then hand it to SkillToolset:

from google.adk.tools.skill_toolset import SkillToolset
from veadk import Agent
from veadk.skills import VeSkillRegistry

registry = VeSkillRegistry(skill_source_id=skill_space_id)
agent = Agent(tools=[SkillToolset(registry=registry)])

This does not download every skill up front: search_skills fetches the remote skill list in real time, and get_skill downloads and loads a specific skill on demand; when the remote version changes, the new version is re-downloaded.

When loading skills, VeSkillRegistry tolerates non-standard frontmatter fields (community extensions) in SKILL.md, such as allowed-tools, triggers, and requires. These fields do not cause loading failures; they are preserved in the skill frontmatter’s metadata.

Harness skills center

Harness can load both Skill Hub skills and AgentKit skills-center spaces. A plain name or slug in skills identifies a Skill Hub skill; the space: prefix identifies a skills-center space ID. For creating, configuring, and deploying the Harness server, see Harness server deployment. The skills center requires Volcengine credentials that can read the space and its object-storage content. For local runs, set VOLCENGINE_ACCESS_KEY and VOLCENGINE_SECRET_KEY, plus VOLCENGINE_SESSION_TOKEN for temporary credentials. On VeFaaS, the bound IAM role can provide temporary credentials instead.

cd my-harness
veadk harness add \
  --skills "data-visualization-cloud,space:ss-example"

The resulting harness.yaml stores the sources as a list:

skills:
  - data-visualization-cloud
  - space:ss-example

Harness downloads and loads the base skills during startup. space:ss-example loads every skill in that space and combines them with Skill Hub skills in one SkillToolset. You can also add a space for one invocation:

veadk harness invoke \
  --name research-agent \
  --message "Prepare a research report" \
  --skills "space:ss-example"

Invocation-level --skills adds skills on top of the base Harness configuration without changing harness.yaml. If any skill cannot be downloaded, lacks a valid SKILL.md, or violates the ADK skill format, Harness stops that load instead of continuing with a partial skill set.

Skill directory layout

A local skill is a directory containing a SKILL.md file, and may include references/, assets/, and scripts/ subdirectories:

skills/
└── kb-skill/
    ├── SKILL.md
    ├── references/
    ├── assets/
    └── scripts/

SKILL.md must declare name and description in frontmatter, and the directory name must match name:

SKILL.md

---
name: kb-skill
description: A skill that queries the knowledge base and composes an answer.
---

Body...

Legacy local entry (deprecated)

VeADK’s old Agent(skills=..., skills_mode="local") local entry still works but is deprecated; migrate to SkillToolset above. The skills_sandbox and aio_sandbox modes are unaffected and used as before:

Mode	Description
`local`	Deprecated. Loads local skills via VeADK’s old path; migrate to ADK `SkillToolset`.
`skills_sandbox`	Skill hosted in a cloud skill space, executed in a sandbox via the `execute_skills` tool.
`aio_sandbox`	All-in-one sandbox mode, for AgentKit-hosted tool runtimes.

Sandbox mode example:

from veadk import Agent
from veadk.tools.builtin_tools.execute_skills import execute_skills

agent = Agent(
    skills=[skill_space_id],
    skills_mode="skills_sandbox",
    tools=[execute_skills],
)

Dynamic skill loading

When enable_dynamic_load_skills=True is set, the agent reloads skills from their configured sources (local directories or cloud skill spaces) before each turn. When new, modified, or deleted skills are detected, the skill list, instruction, and tools are updated automatically without restarting.

enable_dynamic_load_skills can now be used independently of skills: even without an initial skill list, enabling it initializes the skills toolset and registers the refresh callback.

Overview

Get started

Core components

Observability

Extensions

Frontend

Security

Deployment

Local skills

Cloud skill spaces

Harness skills center

Skill directory layout

Legacy local entry (deprecated)

Dynamic skill loading

​Local skills

​Cloud skill spaces

​Harness skills center

​Skill directory layout

​Legacy local entry (deprecated)

​Dynamic skill loading

Local skills

Cloud skill spaces

Harness skills center

Skill directory layout

Legacy local entry (deprecated)

Dynamic skill loading