About agent skills
An agent skill packages knowledge that an agent needs for a specific kind of task, such as how to read the responses from a particular API. You write the skill once and attach it to any agent that needs it. The agent loads the full skill only when the task calls for it, so a skill that isn’t used in a session adds nothing to that session’s token count.
Skills follow the open Agent Skills standard. If you’ve written skills for other agent tools, you can upload the same skills to Itential Platform.
Agent skills are different from Itential AI Skills. Agent skills give FlowAI agents knowledge while they run in Itential Platform. Itential AI Skills are packs of skills for AI coding agents, such as Claude Code, Codex, GitHub Copilot, and Cursor, that help you administer, operate, and build in Itential Platform from those tools. See Itential AI Skills.
Why use skills
Without skills, you put domain knowledge in the prompt of every agent that needs it. That approach has three problems:
- Cost: The agent sends the full prompt to the model on every run, whether or not the task needs that knowledge.
- Duplication: Each agent has its own copy, so you repeat the same work every time you build a new agent.
- Drift: When an API or a response format changes, you have to find and update every copy. Copies you miss go out of date.
A skill solves these problems. It lives in one place, every agent that uses it gets the same content, and the agent loads it only when it’s relevant.
What a skill contains
A skill is a SKILL.md file, with optional supporting files.
You don’t have to start from scratch. Infrastructure Agent Skills is an open-source library of reviewed skills for network and infrastructure automation. You can upload its skills to Platform and adapt them to your own tools and policies.
To learn how to write and upload a skill, see Create a skill.
How agents load skills
An agent doesn’t receive the full content of its skills when a session starts. Instead, it loads each skill in stages:
- When a session starts, the agent receives the name and description of each attached skill.
- When the model decides a skill is relevant to the task, it loads that skill’s body and a list of the skill’s supporting files.
- If the agent needs more context from the skill’s
references/orassets/files, it loads only the files that are relevant.
The agent uses the description to decide when to load a skill, write it carefully. For tips, see Write a description the agent can act on.
If a skill or file fails to load, the session doesn’t stop. The model receives the error and continues without that content. Each load, successful or not, appears in the session’s activity in Agent Sessions. For details, see View skill loads in Agent Sessions.
Skills, prompts, and decorators
Skills work alongside an agent’s prompt and its decorators. Each has a different job.
As a rule of thumb, if you’d paste the same block of text into more than one agent’s prompt, make it a skill instead.
Skill versions
Each time you upload a skill with an existing name, Platform adds a new version of that skill. Versions are numbered, such as v1 and v2, and never change after you upload them. Each version also has a unique version ID.
When you add a skill to an agent, you choose how the agent uses its versions:
- Pin to a specific version: The agent always uses the version you select, even after someone uploads a new one.
- Always use the latest version: The agent uses the most recent version of the skill each time a session starts, including versions uploaded later.
To always use the latest version, select the version marked (latest) in the Version list, and then turn on Use latest and update automatically. Any other choice pins the agent to the version you select.
A new version never changes a session that’s already running. Agent Sessions shows the version ID that each session loaded, so you can always tell which content an agent used.
Skill access
Skills are shared across your Platform installation. They don’t belong to an agent project, and project roles don’t grant access to them. Instead, each skill has its own permissions, which you can grant to users and groups.
If an agent uses a skill you can’t read, the skill appears as locked. You can still run the agent and remove the skill from it, but you can’t see the skill’s content, change the version the agent uses, or add the skill back.
For details, see Set skill permissions.