> ## Documentation Index > Fetch the complete documentation index at: https://developer.opus.com/llms.txt > Use this file to discover all available pages before exploring further. # Opus Agent > The default AI workhorse for LLM-based reasoning, extraction, and generation with typed inputs and outputs. The Opus Agent is your go-to tool for AI-powered automation. Describe what you want in plain English, and it generates a step-by-step blueprint to accomplish the task. You can refine the blueprint, adjust the steps, or convert it to code for full control. ## Key Capabilities Describe what you want to do, and Opus generates a step-by-step blueprint with the inputs and outputs it needs. Edit the generated steps, add new ones, remove what you don't need, or regenerate with updated instructions. The Builder automatically connects outputs from previous tasks to inputs by matching names and types. Convert any blueprint to Python code for full control while keeping the same inputs and outputs. ## How The Blueprint Works When you describe your task in the central description field, Opus Agent generates a blueprint—a series of **Required Steps** that break down your request into smaller actions. Each step has: * **Objective**: What the step aims to accomplish * **Description**: How the step will be carried out After generation, you can: * **Edit steps** manually to fine-tune the logic * **Add new steps** if something is missing * **Delete steps** you don't need * **Regenerate** by adding more instructions and clicking Generate again The blueprint also auto-generates the Input and Output variables based on what your task needs. You can always adjust these manually. ## Operating Modes Opus Agent offers two modes: The fastest way to get started with AI tasks. **How it works:** 1. Describe your goal in natural language 2. The system generates a multi-step blueprint 3. Input and output variables are auto-inferred 4. Tweak the blueprint as many times as you need—your inputs and outputs stay intact based on your described task **Best for:** * Content generation * Data extraction * Summarization and classification * Reliable processing that behaves as expected, even at enterprise scale Accomplish the blueprint's tasks by writing and running code. **How it works:** 1. Start with a blueprint (same as Lite mode) 2. Switch to Advanced mode 3. Click **Build** to generate the code from your blueprint 4. Run the task using the built code **Best for:** * Custom logic not expressible in blueprints * Fine-grained control over execution * Deterministic post-processing * Advanced users who want full flexibility ## Using Opus Agent Click the **+** button or drag an Opus Agent task into the Builder. In the central description field, write what you want the task to do. For example: *"Extract the customer name, invoice date, and total amount from the document."* Click **Generate**. Opus creates a blueprint with Required Steps, each having an Objective and Description. It also auto-generates the Input and Output variables. Review the steps. Edit, add, or delete them as needed. You can also add more instructions and click **Generate** again to update the blueprint. Use **Auto Link** to automatically connect inputs to matching outputs from previous tasks. You can also link them manually. Run a preview with sample data to make sure outputs look right before activating the workflow. ## Best Practices Define your input and output variable types at the start of your design. * Finalize output structures before building downstream logic * Regenerate the blueprint if IO changes materially * Test with representative data in preview mode When downstream logic will branch or serialize data, use JSON mapped to Object variables: * Explicit schemas reduce parsing errors * Structured outputs are easier to validate * Type discipline improves workflow reliability Always preview tasks with representative data before activation: * Check that output types render as expected * Verify variable mappings are correct * Test edge cases (empty inputs, long content) The **Switch to Advanced** button converts your blueprint into an Advanced Task: 1. Click **Switch to Advanced** 2. Click **Build** to generate the full code from your blueprint 3. Run the task using the built code This gives you full control while preserving your inputs and outputs. ## Common Use Cases ### Document Summarization Extract key points from lengthy documents: ``` Input: document_text (Text) Output: summary (Text), key_points (List) Prompt: "Summarize the following document, extracting the main thesis and up to 5 key points. Return as structured JSON." ``` ### Data Extraction Pull structured fields from unstructured text: ``` Input: raw_text (Text) Output: extracted_data (Object) Prompt: "Extract the following fields from the text: name, date, amount, category. Return as JSON matching the output schema exactly." ``` ### Classification Categorize content into predefined buckets: ``` Input: content (Text), categories (List) Output: classification (Text), confidence (Number) Prompt: "Classify the content into one of the provided categories. Return the category name and a confidence score between 0 and 1." ``` ## Related Resources Direct control over LLM selection and prompting. AI-powered review to validate and repair outputs. Custom Python logic for deterministic transformations. Learn workflow design fundamentals.