Skills and knowledge

A Skill is a reusable piece of an agent’s instructions, kept as a versioned row you can edit without touching code. Each skill has a status, a scope, and activation rules that say which requests and tools it applies to. When an agent runs, the platform assembles its prompt from the active skills that match, in priority order. An agent’s charter is itself a skill. When you turn a resolved pursuit into a skill, it is created active and attached to that agent. Draft skills stay out of prompts until activated.

When you’d use it

• Tune an agent’s behavior by editing its charter skill instead of redeploying.
• Add domain knowledge that several agents can share.
• Review what the platform learned from past runs before letting it influence future ones.
• Keep instructions versioned so a change is a tracked edit, not an untracked prompt tweak.

What you manage

Open AI → Learned → All skills (/admin/ai-skills). The list combines session skills and skills your tenant authored. Filter by Kind and Session role, search names, and use Previous or Next to page through results. Platform prompts is available only to platform operators. For each skill you set:

• Status. Active skills are used; drafts are held out until you activate them.
• Scope and activation rules. Which requests and tools the skill applies to. These are read when the Knowledge Plane composes a prompt for a surface. An agent charter and a skill on an agent’s shelf take other paths, and the editor says so on the row itself.
• Body. The instruction text itself.

An agent’s charter is edited on the agent’s own page and saved as a skill behind the scenes.

Good to know

• Only active skills reach a prompt; a draft never does. Captured knowledge waits for your review.
• Creating a skill from a pursuit activates and attaches it. Review the proposed skill before completing that action. Other drafts require activation.
• The charter is a versioned skill, so changing an agent’s behavior is a tracked edit you can trace.
• A skill reaches a model by one of three paths, and they read different settings. The resolver composes by activation rules and priority. An agent charter is loaded directly by id, so neither applies. A skill on an agent’s shelf loads in shelf order, so priority applies and the activation lists do not. Leaving the lists empty means the skill loads on every call: naming an intent or a tool narrows it, so a populated list can mean it never loads.
• Activating a learned draft only helps if it is on an agent’s shelf. A capture is attached to the agent whose work produced it; one attached to nobody stays unreachable however often it is activated, and the editor now says which agents carry it.

How it connects

Skills are content: the words, not the permissions. Which tools an agent may call and whether a write executes belong to the Safety model, and the run itself belongs to the agentic loop.

• Author and activate skills in the agent-designer guides.
• See how an agent’s charter is a skill in Agents.
• Exact shapes: the typed schemas reference lists every field of the skill spec; the store fl_skills (status, scope, kind, activation rules); every term in the glossary.