Skip to content

How-to guide

Create an agent

Create an AI agent capable of interacting with your ontology and knowledge base.

2 min

Start with a mission you can verify: “find orders for customer CLI-001 and cite the records used”. The Console brings configuration, conversation and execution observation together.

Prerequisites

  • Agent Studio is available in your workspace.
  • You may save an agent and start a session.
  • A usable model is configured and test sources are accessible.
  • You know the expected result for at least one question.

1. Open an agent in the Console

Open Agent Studio and the Console. In the top toolbar, click Agents to open the library; select an existing agent or New agent for a draft. Give it a precise name, such as Customer order reader.

The three areas cover configuration, conversation and observation. Visible options depend on workspace modules and permissions.

2. Define the mission and limits

In Identity and instructions, describe the objective, output format and handling of insufficient evidence:

Find orders for the requested customer in accessible sources. Provide order references and evidence. Ask for clarification if the identifier is ambiguous or the period is missing. Do not invent orders or contractual terms.

Instructions guide answers. The platform enforces permissions; the sentence does not.

3. Choose capabilities

Select an available model in Model. Open Execution and choose a reading profile in Capability profile. Select useful records and documents in Context and sources. Enable the required capabilities in Tools.

A Skill supplies reusable instructions, not additional permissions. See Configure tools to distinguish profiles, selections and effective authorization.

4. Save and verify

Save the draft, open Verification and click Verify configuration. Inspect the applied configuration and differences between your selections and available capabilities. Permissions are also checked when a run starts.

Ask a question about CLI-001. Inspect the called tool and returned evidence in the observation panel. Compare the answer with actual instances; convincing wording is insufficient.

Expected result

A saved agent answers a reading question using verifiable evidence or explains why it cannot answer. Before sharing it, test a missing identifier and an out-of-scope request using Test and evaluate.

Did this page help you achieve your goal?