Agent

Chat-style apps where the model can reason, make decisions, and use tools autonomously

Agents are chat-style apps where the model can reason through a task, decide what to do next, and use tools when needed to complete the user’s request.

Use it when you want the model to autonomously decide how to approach a task using available tools, without designing a multi-step workflow. For example, building a data analysis assistant that can fetch live data, generate charts, and summarize findings on its own.

Agents keep up to 500 messages or 2,000 tokens of history per conversation. If either limit is exceeded, the oldest messages will be removed to make room for new ones.

Configure

Write the Prompt

The prompt tells the model what to do, how to respond, and what constraints to follow. For an agent, the prompt also guides how the model reasons through tasks and decides when to use tools, so be specific about the workflow you expect.

Here are some tips for writing effective prompts:

• Define the persona: Describe who the model should act as and the expertise it should draw on.

• Specify the output format: Describe the structure, length, or style you expect.

• Set constraints: Tell the model what to avoid or what rules to follow.

• Guide tool usage: Mention specific tools by name and describe when they should be used.

• Outline the workflow: Break down complex tasks into logical steps the model should follow.

Create Dynamic Prompts with Variables

To adapt the agent to different users or contexts without rewriting the prompt each time, add variables to collect the necessary information upfront.

Variables are placeholders in the prompt—each one appears as an input field that users fill in before the conversation starts, and their values are injected into the prompt at runtime. Users can also update variable values mid-conversation, and the prompt will adjust accordingly.

For example, a data analysis agent might use a domain variable so users can specify which area to focus on:

You are a data analyst specializing in {{domain}}. Help users explore and understand their data.

When asked a question, use available data tools to fetch the relevant information. If the result suits a visual format, generate a chart. Explain your findings in plain language.

Keep responses concise. If a question is ambiguous, ask for clarification before fetching data.

While drafting the prompt, type / > New Variable to quickly insert a named placeholder. You can configure its details in the Variables section later.

Choose the variable type that matches the input you expect:

Short Text

Accepts up to 256 characters. Use it for names, email addresses, titles, or any brief text input that fits on a single line.

Paragraph

Allows long-form text without length restrictions. It gives users a multi-line text area for detailed descriptions.

Select

Displays a dropdown menu with predefined options.

Number

Restricts input to numerical values only—ideal for quantities, ratings, IDs, or any data requiring mathematical processing.

Checkbox

Provides a simple yes/no option. When a user checks the box, the output is true; otherwise, it’s false. Use it for confirmations or any case that requires a binary choice.

API-based Variable

Fetches variable values from an external API at runtime instead of collecting them from users.

Use it when your prompt needs dynamic data from an external source, such as live weather conditions or database records. See API Extension for details.

Label Name is what end users see for each input field.

Generate or Improve the Prompt with AI

If you’re unsure where to start or want to refine the existing prompt, click Generate to let an LLM help you draft it.

Describe what you want from scratch, or reference current_prompt and specify what to improve. For more targeted results, add an example in Ideal Output.

Each generation is saved as a version, so you can experiment and roll back freely.

Extend the Agent with Dify Tools

Add tools to enable the model to interact with external services and APIs for tasks beyond text generation, such as fetching live data, searching the web, or querying databases.

The model decides when and which tools to use based on each query. To guide this more precisely, mention specific tool names in your prompt and describe when they should be used.

You can disable or remove added tools, and modify their configuration. If a tool requires authentication, select an existing credential or create a new one.

To change the default credential, go to Tools or Plugins.

Maximum Iterations

Maximum Iterations in Agent Settings limits how many times the model can repeat its reasoning-and-action cycle (think, call a tool, process the result) for a single request.

Increase this value for complex, multi-step tasks that require multiple tool calls. Higher values increase latency and token costs.

Ground Responses in Your Own Data

To ground the model’s responses in your own data rather than general knowledge, add a knowledge base.

The model evaluates each user query against your knowledge base descriptions and decides whether retrieval is needed—you don’t need to mention knowledge bases in your prompt.

The more detailed your knowledge base description, the better the model can determine relevance, leading to more accurate and targeted retrieval.

Configure App-Level Retrieval Settings

To fine-tune how retrieval results are processed, click Retrieval Setting.

There are two layers of retrieval settings—the knowledge base level and the app level.

Think of them as two consecutive filters: the knowledge base settings determine the initial pool of results, and the app settings further rerank the results or narrow down the pool.

• Rerank Settings

  • Weighted Score

    The relative weight between semantic similarity and keyword matching during reranking. Higher semantic weight favors meaning relevance, while higher keyword weight favors exact matches.

    Weighted Score is available only when all added knowledge bases are indexed with High Quality mode.

  • Rerank Model

    The rerank model to re-score and reorder all the results based on their relevance to the query.

    If any multimodal knowledge bases are added, select a multimodal rerank model (marked with a Vision tag) as well. Otherwise, retrieved images will be excluded from reranking and the final output.

• Top K

  The maximum number of top results to return after reranking.

  When a rerank model is selected, this value will be automatically adjusted based on the model’s maximum input capacity (how much text the model can process at once).

• Score Threshold

  The minimum similarity score for returned results. Results scoring below this threshold are excluded. Use higher thresholds for stricter relevance or lower thresholds to include broader matches.

Search Within Specific Documents

By default, retrieval searches across the entire knowledge base. To restrict retrieval to specific documents, enable manual or automatic metadata filtering.

This improves retrieval precision, especially when your knowledge base is large or contains content for different contexts.

For creating and managing document metadata, see Metadata.

Process Multimodal Inputs

To allow users to upload images, audio, or documents when using the app, select a model that supports the corresponding modalities and enable specific file types—Vision, Audio, and Document.

You can quickly identify a model’s supported modalities by its tags.

Click Settings under Vision to configure how files are accepted and processed. Upload settings apply across all enabled file types.

• Resolution: Controls the detail level for image processing only.

  • High: Better accuracy for complex images but uses more tokens

  • Low: Faster processing with fewer tokens for simple images

• Upload Method: Choose whether users can upload from their device, paste a URL, or both.

• Upload Limit: The maximum number of files a user can upload per message.

Agents also support optional features like conversation openers, follow-up suggestions, text to speech, speech to text, citations and attributions, content moderation, and annotation replies. See App Toolkit for details.

Debug & Preview

In the preview panel on the right, you can test your agent in real time. Select a model, type a message, and send it to see how the agent responds.

We recommend selecting models that are strong at reasoning and natively support tool calling. After selecting a model, you can adjust its parameters to control how it generates responses. Available parameters and presets vary by model.

The system automatically determines the agent mode for the selected model—Function Calling or ReAct. Check the current mode in Agent Settings.

To compare outputs across different models, click Debug as Multiple Models to run up to 4 models simultaneously.

Publish

When you’re happy with the results, click Publish to make your app available. See Publish for the full list of publishing options.