How AI agents work
An AI agent is a system configured to carry out tasks. Depending on its setup, it can plan work, use tools and combine the results of several steps into an answer.
Beyond a question-and-answer conversation, an agent can support tasks with several stages, such as gathering information, comparing it and writing a summary. Its actual capabilities depend on its model, instructions and available tools.
The name “agent” does not automatically mean a solution can browse the web, change files or take action independently. Some marketplace entries are detailed role instructions; others have their own run page. Always check the description of the specific entry.
Goal
Describe the outcome you want, such as “Compare these three supplier quotes.”
Input
Provide the data or text to compare. The system can only use information it can actually access.
Criteria and boundaries
Name the factors that matter, such as price, deadlines, included services and missing information. Specify which actions need your approval first.
Output format
Ask for a specific format, such as a comparison table, a short summary and a list of questions to resolve.
1. Open the runnable agents
Choose a solution that fits your task and read its input requirements. Buying marketplace content and running a solution are separate actions.
2. Set up access
Sign in, enable two-factor authentication and add your own OpenRouter API key. Then select an available model.
3. Enter your task
Give clear instructions and start the run. A task with several stages may make multiple model calls and cost more as a result.
4. Review the answer
Check the figures, conclusions and sources. Supply any missing input, and only use results you have reviewed.
An agent may choose its next step according to its configuration. A workflow follows a predefined sequence, such as creating an outline, drafting the text and checking it. The platform includes examples of both.
Choose according to the task. A workflow can be convenient for recurring work that follows the same sequence; a suitably configured agent can help when the requirements vary.
First check that the solution supports your task. Make the input and output requirements more specific. If necessary, split a large task into smaller parts.
If a run fails, read the error message. Common causes include a missing API key, insufficient provider credit or access, and a temporarily unavailable model. Marketplace credits and your OpenRouter balance are separate.