Agents versus Skills

Agents and Skills both help Glean perform work, but they have different responsibilities. An agent owns an outcome and manages a run. A Skill provides reusable instructions, templates, and task-specific expertise that an agent or Assistant can load when relevant.

Compare agents and Skills

Dimension	Agent	Skill
Ownership	Owns a goal, run, permissions, tools, and output.	Owns reusable guidance for a specific task or domain.
Execution	Starts a run manually or from a trigger, then works toward an outcome.	Is loaded during a response or agent run; it does not start a run by itself.
Orchestration	Chooses or follows steps, tools, flow, memory, and subagents.	Describes how to perform a task and which supporting resources or tools to use.
Lifecycle	Has a draft, test, publish, access, monitoring, and maintenance lifecycle.	Can be created, edited, enabled, shared, imported, and replaced independently.
Best fit	A multi-step workflow that must produce an outcome or take action.	A repeatable method, checklist, format, or body of expertise.
Example	Review a pull request, identify risks, and post an approved summary.	Apply the team's pull request review checklist and output format.

A Skill can be part of an agent. For example, an agent can use a review Skill while it gathers the pull request, chooses tools, and routes the result for approval. The Skill does not replace the agent's trigger, permissions, execution, or publishing controls.

Choose an agent when

• The work needs multiple steps or connected tools.
• The process should start from a schedule, event, or user request.
• The workflow needs branching, memory, subagents, or controlled write actions.
• A team needs to publish, monitor, and govern a shared automation.

Choose a Skill when

• You want the same task-specific guidance in Assistant and multiple agents.
• The main value is expertise, a checklist, a decision method, or an output template.
• The user should decide when to apply the guidance instead of a trigger starting an automated workflow.
• You want to update guidance without owning a separate execution flow.

See Skills to learn how Skills work and create a Skill via chat to get started. To build an agent that uses a Skill, see Create agents with Auto mode.

Use both when the task needs expertise and execution

A common design is to keep the domain guidance in a Skill and the execution in an agent:

1. The agent receives a request or trigger.
2. The agent loads a Skill that defines the required method.
3. The agent uses approved tools and knowledge sources.
4. The agent produces an output or takes an approved action.
5. The agent records the run for testing, monitoring, and improvement.

This separation lets subject-matter experts improve the guidance while builders maintain the agent's execution and governance.