Agents
An agent is a governed AI assistant that answers questions and runs analysis over your Nexa data using the tools you assign it — Genie spaces for natural-language SQL over your tables and metric views, and vector search indexes for retrieval over unstructured knowledge. Each agent is built on a chosen LLM (large language model) serving endpoint, given a system prompt, deployed as a serving endpoint on your data platform, and optionally surfaced as a Streamlit app or a Slack channel. Use the Agents screen to find, monitor, start/stop, and manage agents across the dev, test, and prod environments.
What an agent is made of
Section titled “What an agent is made of”Every agent carries a design record with these core parts:
| Part | What it is |
|---|---|
| Agent name & version | Logical name plus a version (defaults to v1); a generated alias combines them |
| LLM model | The serving endpoint that backs the agent |
| Description & welcome message | What the agent does and the greeting it opens with |
| Orchestration prompt | The system prompt that tells the agent how to route queries between its tools |
| Structured tools | Genie spaces (SQL / natural-language-to-SQL over tables and metric views) |
| Unstructured tools | Vector search indexes (retrieval / RAG over documents and knowledge) |
| Integrations | Optional Streamlit app and Slack channel, plus a workspace access group |
| Runtime settings | Temperature, max tokens, and timeout |
Structured and unstructured tools are the two tool categories an agent can use. See Creating an Agent for how you select them, and Agent Lifecycle for how an agent moves from draft to deployed.
The Agents list
Section titled “The Agents list”The Agents screen lists every agent as a card (grid) or a row (table) — toggle between the two views. Search agents by keyword, and sort by Last Modified, Created Date, Agent Name, Status, or Version. Select Create New Agent to start the create form (available in the dev environment).
Each card shows the agent’s avatar, name, version, status, description, its LLM model, and a Tools count badge whose tooltip breaks the count into structured and unstructured. When a component has been deployed, link icons open it directly:
| Icon | Opens |
|---|---|
| Databricks | The agent playground for the serving endpoint |
| AWS | The Streamlit app |
| Slack | The Slack channel |
The table view adds sortable Agent, Version, Status, Tools, Links, and Created On columns.
Status groups
Section titled “Status groups”An agent’s badge collapses two things — its deployment lifecycle before it is running, and its runtime state after it is deployed. The list groups agents into these buckets, each with a quick-filter tab at the top:
| Group | Tab | What it means |
|---|---|---|
| Started | Started | Deployed and running |
| Deployment Running | In Progress | A deploy or validation is in flight |
| Needs Attention | Needs Attention | A validation or deployment failed and needs action |
| Stopped | Stopped | Deployed but its compute is stopped |
| Draft | Draft | Being authored; not yet deployed |
| Deleted | — | Removed from the environment |
The specific labels you’ll see on a badge include Draft, Validating, Validation Failed, Validated, Artifacts Ready, Artifact Generation Failed, Deployment Running, Deployment Failed, Deployed, Starting, Started, Stopping, Stopped, Needs Attention, and Deleted. A Deployment Running badge shows an inline spinner, and a warning icon with a tooltip appears on agents that need attention.
Actions from the list
Section titled “Actions from the list”Depending on an agent’s state, its card or row exposes:
- Start / Stop — a play/pause control shown when the agent is Started or Stopped.
- View — opens the agent detail screen.
- Traces — opens the agent’s traces (when available).
- Deployment logs — opens the deployment progress panel.
- An overflow menu (dev only) with Edit or Clone & Edit, Promote, and Delete.
Whether you get Edit or Clone & Edit depends on state: a draft or validated agent can be edited in place, while a deployed or failed agent must be cloned into a new version first. Agents mid-flight (validating, deploying, finalizing) are temporarily locked.