Agent Overview
GeniSpace agents combine a model with instructions, authorized information, and tools. The product exposes four agent forms, built on two user-facing runtime patterns: interactive conversations and structured tasks.
Four agent forms
| Form | Runtime pattern | Created by | Best for |
|---|---|---|---|
| Conversational agent | Interactive | Space user | Multi-turn research, support, analysis, and tool-assisted work |
| Task agent | Structured task | Space user | Validated input/output, workflows, batch or event-driven execution |
| GeniAssistant | Interactive | Platform | Managing GeniSpace resources through authorized platform tools |
| Built-in agent or Copilot | Interactive or structured | Platform | A focused capability embedded in Workbench, eBook, Console, or a GeniApp |
Conversational and task agents are resources you configure in Console. GeniAssistant and built-in agents have platform-defined identities and capability boundaries; they do not appear as ordinary editable agent records.
Conversational agents
Use a conversational agent when the user and agent need to exchange information over multiple turns.
Key capabilities include:
- Text, image, and document input when supported by the selected model
- Conversation history and optional memory
- Knowledge-base and dataset retrieval
- Web search, operators, platform tools, and MCP tools
- Progress events, tool results, evidence, and user confirmation
- Automatic and manual context compaction for long sessions
Open the agent in Chat to test the complete interaction.
Task agents
Use a task agent for a defined operation with structured input and output.
- Define input and output schemas.
- Validate values before and after execution.
- Invoke from Console, an API, a workflow, or a Workbench component.
- Use the returned structured value as a downstream workflow input.
Task agents do not use the interactive Chat confirmation flow. Required inputs should be expressed in the input schema or supplied by the invoking workflow.
GeniAssistant
GeniAssistant is the platform Assistant available from Chat. It can explain the platform and, when used in Agent mode, use authorized Assistant tools to create or manage resources such as agents, data sources, tasks, and workbenches.
The Assistant follows the active user's identity, space, permissions, and billing ownership. It cannot bypass a permission that the user does not have.
Built-in agents and copilots
Built-in agents provide a product-specific experience without requiring a user-created agent record. Examples include:
- Workbench Copilot — edits the current Workbench draft through local tools.
- Knowledge Copilot — answers from authorized knowledge and returns evidence.
- Document and data assistants — perform focused structured operations inside an application.
The hosting application determines which local tools are available and how their results are rendered.
Ask mode and Agent mode
| Mode | Behavior |
|---|---|
| Ask | Produces a direct answer without delegating platform actions or multi-step tool work |
| Agent | Can decide, retrieve information, call authorized tools, request confirmation, and continue until it reaches a supported result or explains why it cannot |
Select Agent mode when the request requires live data, a platform change, or tool execution. A mode does not grant additional permissions; it only changes the allowed interaction behavior.
What the execution timeline shows
The execution timeline can show:
- Retrieval and preparation status
- Decisions such as “execute tools” or “generate response”
- Tool name, supplied arguments, status, duration, and result
- User-input requests
- Outcome and response validation status
It is an operational record, not a display of private model reasoning. Use it to verify what data and tools supported the answer.
Tools and data
An agent can receive tools from several sources:
- Platform tools for datasets, data sources, tasks, storage, and applications
- Built-in Assistant tools
- Custom operators
- External MCP servers
- Local tools supplied by the hosting application
Tool discovery and execution both enforce the current user, space, and configured resource scope. Read, create, update, and delete permissions should be limited separately where available.
Choosing a dataset tool
- Use structured query for exact filters on known field names and values.
- Use full-text search for literal terms and identifiers.
- Use vector search for concepts, experience, intent, or cross-language similarity.
A Top-K vector result means “nearest records,” not “records that satisfy a hard condition.” Apply structured validation when the request contains mandatory criteria.
Knowledge, memory, and context
- Knowledge bases provide authorized document evidence for the current request.
- Datasets provide business records.
- Memory stores selected durable facts or preferences for later conversations.
- Conversation context is the working set sent to the model for the current session.
These are separate systems. Compacting context does not delete the visible conversation or long-term memory. See Context Management and Memory.
When the agent asks for input
If a required subject, target, or authorization choice is missing, an interactive agent can pause and display a single-choice or multiple-choice prompt. Select an option or enter a custom answer, then choose Continue. The same turn resumes from its saved state.
Do not start a second message to answer a visible confirmation card unless the card is no longer available.
Configure an agent
- Open Console → Agents.
- Select Create Agent.
- Choose Conversational or Task.
- Enter the name, description, model, and system instructions.
- Configure multimodal input, memory, knowledge, tools, and data as required.
- For task agents, define the input and output schemas.
- Save and test with representative requests.
- Review permissions before sharing.
Avoid placing secrets in system instructions. Store credentials in the platform's secure configuration and grant access through the appropriate tool or data-source configuration.
Completion and limitations
An Agent-mode conversation continues only while it can make supported progress within its iteration and tool limits. A final response should report completed work, evidence, missing information, or the reason further progress is not possible. Wording such as “I am searching” does not mean work continues after the final response.