Agent Instructions

Experimental Feature

Agent Instructions is an experimental feature. Its behavior, limits, and user interface may change in future releases.

Agent Instructions let organization administrators add business rules and facts that an AI Assistant agent should follow when answering questions.

Use instructions for guidance such as business definitions, preferred metric combinations, or rules for handling specific requests. For tone and communication style, use Personality instead.

For example, an instruction can define:

  • Title: Sales = total and net sales
  • Content: When asked about sales, show both Total Sales and Total Net Sales.

Instructions belong to a single agent. They are configured together with the agent in AI Hub and do not exist independently.

Add Instructions to an Agent

  1. Open AI Hub.
  2. Open Agents and select the agent you want to configure.
  3. In Instructions, click Add instruction.
  4. Enter a title and instruction content.
  5. Select how the instruction should be applied.
  6. Save the agent.

The instruction becomes part of the agent configuration and is applied when that agent handles a request.

Application Modes

An instruction can use one of the following application modes:

  • Always: the instruction is applied to every request handled by the agent.
  • Automatic: the instruction is applied when it is semantically relevant to the user’s question.

Automatic matching currently uses the instruction title and content. There is no separate keyword configuration.

Instructions and Personality

  • Use Personality to control how the agent communicates, such as tone, level of detail, or writing style. For example: Keep answers concise and business-friendly.
  • Use Instructions for rules and facts the agent should follow when answering. For example: When asked about sales, show both Total Sales and Total Net Sales.

Do not use Personality as a replacement for business rules that should be consistently applied by the agent.

Instructions and AI Memory

Agent Instructions are separate from AI Memory.

An instruction belongs to exactly one agent and does not participate in organization or workspace inheritance. It is applied only when that agent is active.

Instructions cannot currently be shared or reused across multiple agents.

Manage Instructions Through the API

Instructions are part of the agent definition. They are created and updated together with the agent.

For example:

{
  "data": {
    "id": "<AGENT_ID>",
    "type": "agent",
    "attributes": {
      "instructions": [
        {
          "title": "Sales = total and net sales",
          "strategy": "ALWAYS",
          "content": "When asked about sales, show both Total Sales and Total Net Sales.",
          "isDisabled": false
        }
      ]
    }
  }
}

Use:

PUT /api/v1/entities/agents/<AGENT_ID>

Both content and strategy are required for each instruction.

Important Notice

Updating an agent replaces its complete instruction list. If the update omits instructions, the existing instructions are removed. When updating instructions through the API, send the complete list that you want the agent to keep.

Current Limitations

The following limitations apply to the experimental version:

  • An instruction belongs to one agent only. Sharing instructions across agents is not supported.
  • Workspace-level and organization-level instructions are not supported.
  • Each instruction can contain up to 255 characters of content.
  • An agent can contain up to 50 instructions.
  • Automatic matching uses only the instruction title and content. Separate matching keywords are not supported.
  • Instructions cannot currently be created through an MCP tool.