--- title: "Agents and Skills: What the Difference Is" url: https://www.insulin.dev/blog/agents-and-skills/ canonical: https://www.insulin.dev/blog/agents-and-skills/ type: Blog description: "An agent is a configured assistant; a skill is an instruction package it loads. What each is for in Insulin, and which one your problem actually needs." --- # Agents and Skills: What the Difference Is > Canonical HTML version: https://www.insulin.dev/blog/agents-and-skills/ 1. [Home](/) 2. / 3. [Blog](/blog/) 4. / 5. Agents and Skills: What the Difference Is # Agents and Skills: What the Difference Is An agent is who you talk to. A skill is what it knows how to do. Getting the distinction right decides whether you build one thing or twenty. ![Gabriel Paiva](/authors/gabriel-paiva.jpg) Gabriel Paiva Product Lead · Aug 16, 2026 ![Agents and Skills: What the Difference Is](/images/blog/agents-and-skills/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fagents-and-skills%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Agents%2C%20Agent%20Marketplace. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fagents-and-skills%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Agents%2C%20Agent%20Marketplace. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fagents-and-skills%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Agents%2C%20Agent%20Marketplace. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fagents-and-skills%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Agents%2C%20Agent%20Marketplace. "Summarize with Perplexity") Table of Contents - [What is an agent?](#what-is-an-agent) - [What is a skill?](#what-is-a-skill) - [The test that decides which you need](#the-test-that-decides-which-you-need) - [Where most teams get it wrong](#where-most-teams-get-it-wrong) - [A shape that scales](#a-shape-that-scales) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _An agent is a configured assistant you have a conversation with. A skill is a package of instructions an agent loads while working. The difference matters because most teams’ first instinct — build a new agent for every task — produces a list nobody can navigate within a month._ * * * The first week with a workspace like this, everybody builds agents. One for the weekly report. One for the pricing questions. One for the thing that happened last Tuesday. By week four there are thirty of them, four are used, and nobody can remember which of the remaining twenty-six does the invoice thing. The fix is not discipline. It is knowing that there are two different objects here, and that most of what people build as agents should have been skills. * * * ## **What is an agent?** **An agent is “an AI assistant with defined instructions, model configuration, optional integration access, optional skills, and optional knowledge bases.”** It is the thing you address. It has a name, an icon, and a personality defined by its system prompt. Configuring one means choosing: - **Ownership** — personal or organisational - **Default model** — from your connected providers - **Knowledge bases** — optional, attached in Read or Edit mode - **Integrations** — an optional allowlist of connected systems - **System prompt** — the behavioural instructions There is also a built-in agent, described as “your chief of staff: a general-purpose assistant.” It cannot be edited or deleted, which is a feature — there is always one that works. * * * ## **What is a skill?** **A skill is “a markdown-based instruction package containing procedures or domain knowledge”** that agents load during conversations to inform their responses. It has no personality, no model, no integrations. It is knowledge and procedure in a form an agent can pick up. Think of it as the difference between hiring a person and handing them a runbook. * * * ## **The test that decides which you need** Ask one question: **does this need its own conversation, or does it need to be available inside conversations you are already having?** Signal Build Distinct audience or purpose Agent Needs its own integrations or model Agent Should have a different personality or tone Agent A procedure to follow Skill Domain knowledge several agents need Skill A format or template to apply Skill The strongest signal for a skill is **reuse**. If two agents would both need to know it, it is a skill. Duplicating a procedure into two system prompts guarantees that in six months they say different things and nobody knows which is current. The strongest signal for an agent is **a different set of hands**. If the work needs integrations another agent should not have, that access boundary is the agent boundary. Organisation-level agents can access only org-level integrations and cannot use user-level ones, so scope is not something you can fudge after the fact. * * * ## **Where most teams get it wrong** **Building an agent per task.** A task is almost always a skill, or a [job](/jobs/) if it should run without being asked. Agents multiply badly because each one is a thing to find, name, and maintain. **Putting procedure in the system prompt.** A system prompt should say who the agent is and how it behaves. The steps for handling a renewal are a skill — extractable, reusable, and editable without touching the agent’s character. **Building before browsing.** Both agents and skills can be installed from the [Agent Marketplace](/agent-marketplace/), and installing is idempotent — re-installing the same item does not create a duplicate. After installation they “behave identically to ones you create yourself,” so an install is a starting point rather than a commitment. [Whether to build or install](/blog/reusable-ai-agents-build-or-install/) covers that choice in full. * * * ## **A shape that scales** Few agents, many skills. Two or three agents matching the real divisions in your team — the ones with genuinely different data access — and a growing library of skills they share. New capability usually means a new skill attached to an agent that already exists, which is a smaller change and a smaller thing to explain. When a skill grows to the point that it needs its own integrations or its own audience, promote it to an agent. That is a real signal, and it arrives rarely. [One assistant or many specialists](/blog/one-ai-assistant-or-many-specialists/) covers where to draw that line. * * * ## **Frequently asked questions** **What is the difference between an agent and a skill?** An agent is a configured assistant with instructions, a model, and optional integrations and knowledge bases. A skill is a markdown instruction package containing procedures or domain knowledge that agents load while working. **Should a new task be an agent or a skill?** Usually a skill. Build an agent when the work needs its own audience, its own integrations, or a genuinely different behaviour. **Can several agents use the same skill?** Yes, and that is the main reason to make something a skill. Shared procedures stay in one place instead of being copied into multiple system prompts. **Can an installed agent be changed afterwards?** Yes. After installation, agents and skills from the marketplace behave identically to ones you create yourself, so prompts and configuration can be edited. **Does installing twice create duplicates?** No. Installing is idempotent — re-installing the same item does not create a duplicate copy. **Can an org agent use my personal integrations?** No. Organisation-level agents can access org-level integrations only, and cannot use user-level ones. * * * ## **Takeaways** - An agent is who you talk to; a skill is what it loads while working. - Reuse is the signal for a skill. If two agents would both need it, do not copy it into two prompts. - A different set of integrations is the signal for an agent, because access scope cannot be fudged later. - Procedure belongs in a skill, not in a system prompt. Character belongs in the prompt. - Aim for few agents and many skills. Promote a skill only when it needs its own audience or access. - Browse before building — installs are idempotent and fully editable afterwards. * * * Most of what teams build twice should have been built once. See what is already in the [Agent Marketplace](/agent-marketplace/) before you write another system prompt. ## Sources Primary sources for the platform rules cited above. Last verified August 16, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Suger Insulin docs: Marketplace](https://doc.suger.io/insulin/marketplace/) — The definitions of agent and skill quoted verbatim, that a skill is a markdown instruction package agents load during conversations, and that installed items behave identically to ones you create - [Suger Insulin docs: Agents](https://doc.suger.io/insulin/agents/) — That an agent is an AI assistant with instructions, model configuration, optional integration access, optional skills and optional knowledge bases, and the org-level versus user-level integration scope rules ### Stay Updated Get the latest Cloud GTM insights, product updates, and marketplace strategies delivered to your inbox.