Agent vs Subagent vs Skill vs Plugin vs Connector: What's the Difference in AI Tools in 2026

If you’ve spent any time inside Claude, ChatGPT, Microsoft Copilot, or one of the no-code automation tools lately, you’ve probably hit all five of these words in the same week: agent, subagent, skill, plugin, connector. They sound related. They get used in overlapping sentences. And somehow you still can’t tell whether a “plugin” and a “connector” are the same thing wearing different vendor branding.

They’re not interchangeable, but they’re also not used the same way by every company. That second part is the real source of the confusion, and it’s worth saying plainly up front: there is no single industry standard here. Anthropic, OpenAI, Microsoft, and the no-code platforms each settled on their own vocabulary, and those vocabularies don’t map onto each other perfectly. This article won’t pretend otherwise. Instead, it walks through what each term tends to mean, where vendors disagree, and how the five concepts typically fit together once you’ve seen them in practice.

Why the Confusion Happens

Part of the problem is timing. These features didn’t arrive together as a tidy, pre-planned system, they rolled out in waves, often a year or more apart, as companies kept extending the same core product. Anthropic shipped subagents, then Skills, then Connectors, then Plugins, each solving a different problem that came up after the last one shipped. OpenAI built custom GPTs, then plugins, then renamed “connectors” to “apps,” then introduced Workspace Agents as a replacement for GPTs. Microsoft has agents, plugins, and connectors too, but draws the lines between them differently than either Claude or ChatGPT does.

The other part of the problem is that these words were already common English before AI vendors picked them up. “Agent,” “skill,” and “plugin” all carried everyday meaning before any of this software existed, so it’s easy to assume you already know what they mean in a new context, and easy to be wrong in a way that doesn’t feel wrong.

With that caveat in place, here’s what each term tends to describe.

Agent

An agent, in the current AI sense, is a system built around a language model that can take a goal, break it into steps, use tools to act on those steps, check its own results, and keep going until the task is done, without a human approving every single step along the way.

That’s different from a plain chatbot. A chatbot answers what you type. An agent can decide on its own that it needs to search the web, read a file, run a calculation, or send a message, and then actually do those things before getting back to you.

A concrete example: ask an AI chatbot “what’s our refund policy?” and it answers from what it already knows. Ask an agent “find any customers who requested a refund last week and draft a reply to each one,” and the agent will go look up the data, reason about which customers match, draft the replies, and present them to you, often across several internal steps you never see directly.

Not every vendor’s “agent” works at the same scale. OpenAI’s Workspace Agents run continuously in the cloud and connect to tools like Slack and Salesforce around the clock. Claude’s agentic mode inside a single conversation is more contained, it acts within that session unless you’ve set up something more persistent. Both are accurately called agents; they just operate at very different scopes.

Subagent

A subagent is a separate, smaller agent that a main agent calls on to handle one specific piece of a larger job, then reports back to the main agent with a summary. The defining trait is isolation: a subagent runs in its own separate workspace and context, so whatever it does, reading through a hundred files, running a long search, drafting a section of a document, doesn’t clutter or distract the main conversation. It hands back a clean result.

Think of it like delegating a task to a coworker instead of doing it yourself in the middle of a meeting. You don’t need to see every step they took; you just need the outcome.

A practical example, drawn from how Claude Code uses the term: you might ask a main coding agent to refactor a large codebase. Instead of doing every step itself, it can spin up a subagent specifically to research how a particular library is used across the project, get back a concise report, and then continue the refactor with that information in hand, without the research process eating up the main agent’s working memory.

The relationship is hierarchical and one-directional. A subagent doesn’t talk to other subagents, and it doesn’t make decisions about the overall project. It does its assigned piece and stops.

Skill

A skill is a set of instructions, usually a folder containing a written guide, and sometimes scripts or reference files, that teaches an agent how to do a specific task the way you (or your organization) want it done. It’s less like installing new software and more like handing the agent a playbook for one job.

The defining feature of how skills work in most current systems is that they load on demand rather than sitting active all the time. The agent first sees just the name and a short description of each available skill. Only when a task actually matches that description does the agent pull the full instructions into its working context and follow them. This keeps the agent fast and uncluttered most of the time, while still giving it access to a large library of specialized know-how when it’s actually needed.

A concrete example: a skill might contain step-by-step instructions for how your company formats a weekly status report, which sections to include, what tone to use, where to pull the numbers from. When you ask an agent to “write this week’s status report,” it recognizes the match, loads that skill’s instructions, and follows them, rather than guessing at a format from scratch.

The Agent Skills format started as an Anthropic-specific feature but was released as an open standard, and it’s since been picked up by other agent products beyond Claude, which makes “skill” one of the more portable terms on this list, at least in its newer, more technical sense. That said, plenty of platforms also use “skill” loosely to mean any specialized capability, including in marketing copy that has nothing to do with the formal Agent Skills format. Read the word in context.

Plugin

A plugin is a packaged bundle that combines several other pieces, typically some mix of skills, connectors, commands, and sometimes subagents, into one installable unit aimed at a specific job or workflow. The point of a plugin is distribution and convenience: instead of manually assembling five separate pieces every time you want a certain capability, you install one plugin and get the whole set.

This is one of the spots where vendors diverge the most. In Claude Code, a plugin can bundle skills, hooks (automated triggers tied to events like a file save), subagents, and MCP servers, connections to outside services, into a single package you install from a marketplace with one command. In ChatGPT’s earlier terminology, “plugins” referred to something narrower: a way to give the chatbot access to a single external service or API, more like what Claude now calls a connector. OpenAI has since moved away from that usage, replacing plugins-as-a-concept with “apps” and, for business accounts, “Workspace Agents.” Microsoft’s plugins sit closer to OpenAI’s older definition, a plugin there is a component that gives an agent access to a specific external capability via an API, which is a narrower job than what a Claude Code plugin does.

So when someone says “I installed a plugin,” the honest answer to “what did that actually do?” is: it depends which product they mean.

Connector

A connector is the piece that gives an agent access to a specific outside service or data source, Gmail, Slack, a database, Salesforce, Google Drive, so the agent can read from it or act on it directly, instead of you having to copy and paste information back and forth manually.

Most current connectors are built on the Model Context Protocol (MCP), an open standard originally released by Anthropic that defines a common way for an AI system to talk to outside tools and services. Because it’s an open standard, a connector built for one MCP-compatible product can often work with others, which is part of why the term has spread across vendors faster than some of the others on this list.

A concrete example: connecting Claude to your Linear account lets it create or update issues directly, using your existing permissions, it can only see and do what you could already see and do logged in yourself. Microsoft draws a further distinction here that’s worth knowing: it splits connectors into “synced” connectors, which index outside data into its own search system in advance, and “federated” connectors, which use an MCP-style approach to fetch data live, without indexing it first. That’s a meaningful technical difference if you’re choosing between them, even though both get called “connectors.”

Worth flagging again: this is also the term that changed names recently at OpenAI. What ChatGPT called “connectors” became “apps” in December 2025, partly to unify the idea of connectors (which pull in your data) with richer in-chat tools that have their own interface. If you’re reading an older article that mentions “ChatGPT connectors,” it’s describing what’s now called an app.

How They Compare

Term	What it is	Persists across sessions?	How it’s invoked	Typical scope
Agent	A goal-driven system that plans, acts, and self-corrects using tools	Depends on setup, can be a single session or run continuously	You give it a goal in plain language	Can span an entire task or workflow
Subagent	A separate agent instance spawned by a main agent for one scoped piece of work	No, exists for the duration of its assigned task, then returns a result	Called automatically by the parent agent, not directly by the user	Narrow, one delegated task
Skill	A set of instructions (and sometimes files or scripts) for doing one task a specific way	Yes, saved and reusable, but only loaded into context when relevant	Auto-matched to a task, or triggered manually (e.g., a slash command)	One specific task or format
Plugin	A bundle that packages skills, connectors, commands, or subagents together	Yes, installed once, available afterward	Installed from a marketplace or repository; runs the components inside it	A full job or workflow, made of several pieces
Connector	A live link between an agent and one outside service or data source	Yes, set up once, stays connected until removed	The agent calls it mid-conversation when it needs that service	One external system at a time

How They Fit Together

None of these exist in isolation, and seeing how they nest is often what makes the rest click.

A running agent is the thing actually doing the work. While it works, it can load a skill, pulling a specific set of instructions into its context so it follows your preferred process for that task. If the agent needs to delegate a self-contained chunk of work without cluttering its own context, it spins up a subagent, which works independently and reports back a result. If the agent needs to reach an outside service, your email, your project tracker, a database, it uses a connector to do that, acting within whatever permissions you already have there. And a plugin is the packaging layer above all of this: it’s how a skill, a connector, and sometimes a subagent get bundled together and distributed as one installable unit, so you don’t have to set each piece up separately.

A rough way to hold this in your head: the agent is the worker, a skill is a page from its instruction manual, a subagent is a temporary helper it brings in for one task, a connector is a key to a specific door, and a plugin is the toolbox that showed up with several of those keys and manuals already inside it.

It’s also fine to notice that this mental model fits Claude’s current product lineup more cleanly than it fits every vendor. Microsoft’s connectors split into two technically different categories. OpenAI folded its connectors into a broader “apps” concept and is phasing out the GPT format in favor of Workspace Agents. No-code platforms like Zapier and Make use “integrations,” “modules,” and “agents” in ways that overlap with, but don’t exactly mirror, any of the above, we walk through a concrete Make/Zapier example in how to automate client invoices with AI, if you want to see this kind of agent-plus-connector setup in practice. If a product you’re using defines one of these terms differently than this article does, that’s not necessarily a mistake on their part, it’s the actual, unsettled state of the terminology in 2026. The more useful habit is to ask, for any specific product, “does this run independently, does it persist, and what does it actually connect to or instruct?” Those three questions get you to the real answer faster than the label does.

If you’re less interested in the terminology itself and more in which tools to actually use day to day, our 7 best AI tools for freelancers in 2026 roundup covers the practical side, writing, video, voice, and editing tools, without needing to know whether what you’re using is technically an agent, a plugin, or a connector.