Creating an Agent

The Create New Agent form walks you through configuring an agent, validating it, and deploying it in a single flow. Use it when you want a new AI assistant over your Nexa data. Creating agents is a dev-environment action — you author and deploy to dev first, then promote to test and prod. Open it with Create New Agent on the Agents screen.

The form has three steps, shown as pills across the top: Configuration, Validation & Artifacts, and Deployment.

Step 1 — Configuration

This step defines the agent’s identity, model, tools, and integrations.

Agent Basic Information

1. Agent Name (required) — up to 25 characters, using letters, numbers, underscores, or hyphens. It must be unique. The read-only Agent Alias below shows the auto-generated slug plus version.

2. LLM Model (required) — search and select the serving endpoint that backs the agent. The options are loaded from your workspace’s available models.

3. Advance LLM Settings (optional) — expand to tune runtime behavior:

   • Temperature — 0.0–1.0 (default around 0.7).
   • Max Tokens — 100–3000.
   • Timeout — 10–300 seconds.

4. Welcome Message (optional) — the greeting the agent opens with.

5. Agent Logo (optional) — an image (JPEG, PNG, WebP, or JPG).

6. Agent Description (required) — describe what the agent should do, minimum 15 characters. This description drives tool recommendations, so edit it to refine which tools Nexa suggests.

Tool selection

Agents use two categories of tools. Select Generate Tool Recommendations to have Nexa suggest tools from your description, then pick from two columns:

Column	Tool type	What it provides
Structured Tools	Structured	Genie spaces — natural-language SQL over your tables and metric views
Unstructured Tools	Unstructured	Vector search indexes — retrieval over documents and knowledge

Each column has a Show recommended / Show all toggle and a Filter tools by name… box. Every tool is a checkbox with its type badge, an optional Recommended badge, and an info tooltip showing the recommendation’s relevance percentage and reasoning. Select the tools the agent should be able to call. At least one structured and one unstructured tool is expected; validation warns you if none are selected or if a tool appears in both columns.

Actions & Integrations

Configure how the agent is surfaced and who can use it:

Control	Effect
Slack Endpoint	When on, auto-creates a dedicated Slack channel for this agent version using the environment’s default Slack app configuration
Streamlit	When on, deploys a Streamlit app front end for the agent
Workspace Access Group	The group granted app access and query permission on the serving endpoint (defaults to users)

When you’re done, select Validate. Nexa saves the draft and runs validation; if it fails, the button changes to Edit Configuration so you can fix and retry. See Agent Lifecycle for what validation checks.

Tip

Slack integration relies on the environment’s Slack app configuration. Set that up first under Slack Integration.

Step 2 — Validation & Artifacts

This step is a read-only review of the whole agent before you commit it. It groups the configuration into:

• Agent Information — name, LLM model, welcome message, logo, description, and the selected structured/unstructured tools.
• LLM Runtime Settings — version, temperature, max tokens, and timeout.
• Actions & Integrations — the Slack and Streamlit toggles.
• Validate Response — the validation status with Errors and Warnings counts, a message, and expandable error and warning details.

Review the summary, then select Create Agent (labeled Create New Agent Version when you’re versioning an existing agent). This commits the agent’s artifacts and dispatches its deployment to the dev environment in one step.

Step 3 — Deployment

This step tracks the deploy you just started. If the agent hasn’t been created yet, it prompts you to go back to the review step and select Create Agent. Once dispatched, you can watch progress two ways:

• Summary — a component-by-component progress view (agent endpoint, Streamlit app, Slack channel) with retry controls.
• Workflow Details — the underlying deployment workflow logs, with a link to the run and a Go to Agents list button.

A confirmation appears when the agent is committed and deployment is dispatched. From here the agent shows up in the Agents list, where you can start/stop it, view its detail, and later promote it to other environments — see Agent Lifecycle.

Editing and cloning

Opening an existing agent in the form loads its current configuration. If the agent is already past validation, the form jumps straight to the review step. A draft or validated agent can be edited in place; a deployed agent is cloned into a new version (Clone & Edit) so the running version stays untouched. When cloning, Nexa can show a diff between the parent version and the new one.