Last updated:
Create agent
In AI Studio, click Create Agent and assign it a name. This name appears under General → Properties in the right-hand sidebar, where the rest of the agent is configured.
General
Section titled “General”| Field | Description |
|---|---|
| Id | The unique identifier of the agent. |
| Name | The name of the agent. Visible in the agent list: should clearly describe its purpose. |
| Note | Optional internal description. Helps explain what the agent does; not used in AI processing. |
| Provider | Which AI provider is used, for example OpenAI GPT-5 or AzureAI GPT-5.2. |
| Agent info | An explanatory display text for the agent. |
When creating an agent you select an agent type, which determines the AI provider (OpenAI, Azure AI, Claude, or Mistral), the model or deployment, and the available properties and behavior of the agent.
- For OpenAI, the selected agent type directly determines the model.
- For Azure AI, additional configuration is required: the selected agent type must match your configured deployment.
- For Claude and Mistral, you pick the model from a list on the agent itself.
OpenAI
Section titled “OpenAI”OpenAI uses direct model selection:
- Connection: your OpenAI API key connection.
- Model: choose a model directly. GPT-5 models are recommended for advanced use cases and production agents; transcription models are for audio-to-text workflows.
Azure AI
Section titled “Azure AI”Azure uses a resource name and deployment instead of a direct model name:
- Connection: your Azure API key connection.
- Resource name: the name of your Azure AI resource, found in the Azure portal.
- Deployment name: the name of the deployed model in Azure. This is not the model name: it’s the deployment name you created in Azure.
Common issues: an incorrect resource name, an incorrect deployment name, or using a model name instead of a deployment name. If the agent doesn’t respond, verify these values first.
Claude
Section titled “Claude”Claude uses direct model selection:
- Connection: your Anthropic API key connection.
- Model: choose Claude Fable 5.1, Opus 5.5, Opus 5, Sonnet 5 (default), or Haiku 4.5. See Claude for guidance.
Mistral
Section titled “Mistral”Mistral uses direct model selection:
- Connection: your Mistral API key connection.
- Model: choose Mistral Medium (default), Large, or Small. See Mistral for guidance.
Settings
Section titled “Settings”Depending on the selected model, additional configuration options may be available.
| Property | Description |
|---|---|
| Connection | The AI connection used by the agent. Must be configured before the agent can execute. |
| Model | The model family, for example GPT-4 or GPT-4 mini. |
| Is chat | Enables chat-based interaction. When disabled, the agent only runs in execution mode. |
| Verbosity | The level of detail the final response should have. |
| Reasoning effort | How much effort the model applies to reasoning. Higher levels generally mean longer requests and more tokens used. Called Effort for Claude (not available for Haiku 4.5), and offers only None or High for Mistral (not available for Mistral Large). |
| Image analysis | The level of detail used when processing and understanding images. |
| Max tool calls | The maximum number of total calls to built-in tools across a single response: not per individual tool. |
| Max file size | The maximum combined size (MB) of files handled in a request. |
| Enable history | Keeps track of previous interactions and enables the history settings below. |
History
Section titled “History”History and file retention settings control how conversations and their file references are stored and cleaned up. Any2Info distinguishes between its own conversation storage and the AI provider’s (temporary or persistent) storage.
Chat enabled: provider calls are stateful, previous messages are included in subsequent requests, and a conversation is always created and stored regardless of the history setting (though it’s only accessible to the user when history is enabled).
Chat disabled: provider calls are stateless, and a conversation is only stored when history is enabled. Files are cleaned up immediately after each request.
File handling: within Any2Info, images are stored as full base64 content, while other file types (PDF, Word, Excel, TXT, …) only keep their file name and extension: content isn’t stored. At the AI provider, images are never stored (sent as base64), and other files are either uploaded to provider storage if supported, or converted to text and sent inline.
Cleanup runs hourly, based on two optional settings:
| Setting | Description |
|---|---|
| Max duration (days) | How long a conversation may stay unused before cleanup. The timer resets whenever it’s used again. |
| Max count | The maximum number of stored conversations per agent. The oldest is removed once the limit is exceeded. |
When cleaned up, a conversation is soft-deleted in Any2Info (data stays in the database) while its files are permanently removed from the AI provider. A user can also trigger the same cleanup manually, for example with the clear messages action.
Instructions
Section titled “Instructions”The system instruction defines the role, behavior, and boundaries of the agent. Without a clear instruction, an agent may behave unpredictably or produce inconsistent results.
Guidelines for writing a good instruction:
- Be explicit about what the agent is and isn’t allowed to do.
- Clearly describe its role (assistant, analyst, validator, …).
- Define input expectations where applicable.
- Avoid unnecessary verbosity, and leave implementation details to dataflows.
A thorough instruction typically covers several areas: the agent’s role and objective, ground rules for interpreting input, how it should interact with the user (tone, avoiding technical jargon, asking for confirmation before running a tool), how it’s allowed to use dataflows, dataclips, files, and vector stores as tools, and the expected output format. How much of this is needed depends on the agent: a conversational chat agent mainly needs guidance on tone and interaction style, while a structured agent (analysis, extraction, task execution) needs stricter rules and output formats.
Inputs are configured as parameters: structured placeholders that can be populated from chats, dataflows, triggers, or external integrations. They’re available both in the system instruction and at runtime.
Supported types: Text, Boolean, Integer, Fraction, Date, Datetime, File, MultipleFiles.
| Field | Description |
|---|---|
| Name | Unique identifier, used to reference the input in instructions and at runtime. |
| Type | The data type of the input. |
| Description | Optional explanation of the parameter’s purpose. |
| Is main chat input | Marks this parameter as the primary chat input. Only one parameter can have this status. |
Output
Section titled “Output”Outputs are also configured as parameters, describing the results an agent is expected to produce. They’re available in the system instruction and runtime context, and can be used as event parameters in triggers, dataflows, or external integrations.
Supported types: Text, Boolean, Integer, Fraction, Date, Datetime.
Important limitation: an agent can’t generate files as direct output: output parameters of type File or MultipleFiles aren’t supported, and using them as event parameters to start a dataflow won’t work. To generate a document such as a PDF, return structured text output and trigger a dedicated dataflow designed for that purpose.
Sources
Section titled “Sources”Sources give the agent read-only, contextual data that’s automatically added during execution. They augment input parameters rather than replace them.
| Source type | Description |
|---|---|
| Dataclips | Datasets created within Any2Info: answers “what exists in my systems?” (e.g. relationships, articles, projects). Used for fuzzy matching, finding IDs, and validating input. |
| Dataflows | Actions within Any2Info (create, update, sync, export, trigger): answers “what is the agent allowed to do?” |
| Files | Uploaded documents (PDF, Word, Excel, images, scans): the primary source the agent reads and interprets, e.g. quotations, invoices, inspection reports. |
| Vector stores | An indexed knowledge base built on embeddings: gives the agent semantic memory for retrieval-augmented answers, policy enforcement, and consistent explanations. |
Tool calling is how an agent actively invokes dataclips, dataflows, file search, or web search during a conversation, instead of just generating text. Clear descriptions on each dataclip, dataflow, and event parameter are critical: the agent relies on them to decide when and how to use a tool, and how to map its inputs and outputs.
Web search
Section titled “Web search”Coming soon.
External MCP tools
Section titled “External MCP tools”An agent can connect to external MCP (Model Context Protocol) servers to use tools they expose: this is the reverse of Any2Info’s own MCP feature, where Any2Info is the one exposing dataflows and dataclips to external clients.
To use an external MCP server, first create a connection for it in Data Studio, then define the MCP server with:
| Property | Description |
|---|---|
| Name | Display name of the MCP server. |
| Url | URL endpoint of the MCP server. |
| Connection | The connection used for authentication and communication. |
Once defined, the MCP server can be selected from the MCP servers dropdown in the agent’s properties panel, and reused across multiple agents.
Only connect to trusted MCP endpoints: data may be shared with the external system behind them.