Skills: Reusable Procedures and Tool Access for Agents

We’ve added Skills to Siesta AI, a dedicated way to define what an agent can do, how it should do it, and which tools it’s allowed to use. If your system prompt answers “who the agent is”, skills answer “what the agent can reliably execute”.

This matters once you move beyond a single demo agent. In real teams, the same request starts producing different outcomes depending on which agent someone uses, because instructions get copied, tweaked, and slowly drift. Skills are designed to stop that drift by turning repeatable work into a maintained capability you can reuse across agents.

Skills in practice: reusable procedures and tool access

A skill is a packaged unit of execution that combines two things: instructions (the procedure) and tools (the functions and integrations the agent is allowed to use for that procedure). When you attach a skill to an agent, you’re doing more than “adding a feature”. You’re setting boundaries that make behavior more predictable.

The non-obvious benefit is governance-through-design. When a workflow is wrapped in a skill, you can keep the agent focused on the few tools it actually needs. That reduces accidental tool use, makes failures easier to debug, and keeps outcomes consistent across teams, even as you add more agents.

Organization skills vs. system skills

Skills are split into two categories. Organization skills are created inside your org. You can create, edit, and delete them. This is where you capture how your company actually works, for example how you want tickets written, how reporting should be formatted, what checks must happen before an email is sent, or which system is the source of truth when data conflicts. System skills are predefined and available to everyone. They’re not editable directly, but you can copy them and adapt the copy. Practically, this gives you a starting point you can tighten to your internal rules, naming conventions, approvals, and allowed tools.

Skills vs. Subagents: when to use which

Skills and Subagents solve different problems, and confusing them leads to messy agent setups.

A skill is about standardizing execution. Use it when you want the same task done the same way every time: the same steps, the same tool access, the same output shape. Skills are your repeatability layer.
A subagent is about delegation and specialization. Use it when a request is too broad or multi-disciplinary for one agent to handle cleanly, so the main agent hands off parts of the work to specialized agents.

The practical pattern that works well is combining them: subagents split the work, skills keep each part consistent. Without skills, subagents can still produce variable outcomes because each specialized agent “does it their own way”. Without subagents, skills tend to become bloated because one agent tries to handle every step of a complex request.

Read more about Subagents: https://siesta.ai/blog/1329/new-in-siesta-ai-subagents

A better way to roll this out

If you want Skills to pay off quickly, don’t start with a “master skill” that tries to cover everything. Pick one workflow that already has clear ownership and repeats weekly, like creating Jira tickets from support escalations, turning meeting notes into follow-ups, or producing a standard ops update.

Write the skill like an internal runbook, give it only the tools it needs, then assign it to the agents people already use. After a week, you’ll know exactly what to tighten: missing inputs, unclear handoffs, permission edge cases, and the two or three failure modes that cause the most rework. Fix those once in the skill, and the improvement propagates everywhere automatically.