Skip to content

Agent Lifecycle

An agent progresses through a defined lifecycle before it can serve traffic: you author it, validate it, generate its artifacts, and deploy it — first to dev, then promote it to test and prod. This page explains each stage, the environments involved, and the runtime controls (start, stop, promote, logs). It complements Creating an Agent, which covers the create form itself.

Deployment moves through these stages, reflected in the agent’s status badge on the Agents list:

Stage Status Meaning
Draft draft Being authored; not yet validated
Validating validating Validation in progress
Validated validated Passed validation; ready to generate artifacts
Validation failed validation_failed Validation found blocking issues
Generating artifacts generating_artifacts Config files are being generated and committed
Artifacts ready artifacts_generated Artifacts committed to git; ready to deploy
Deploying deploying The deployment workflow is running
Deployed deployed The serving endpoint (and any app/Slack) is live
Deployment failed deployment_failed The deployment workflow failed
Deleted deleted Removed from the environment

Once deployed, a second, runtime dimension applies — Starting, Started, Stopping, or Stopped — reflecting whether the agent’s compute is running.

Validation is a required pre-deploy check. From the create form select Validate; Nexa then:

  1. Checks required fields — the LLM model, description, and orchestration prompt must be present, and temperature (0–2), max tokens, and timeout must be valid. It warns if no tools are selected or if a tool is in both the structured and unstructured lists.

  2. Validates the bundle — Nexa validates the generated deployment bundle against your workspace without committing anything to git or deploying.

  3. Records the result — on success the agent becomes Validated; on failure it becomes Validation Failed with error and warning details you can review in the Validation & Artifacts step.

After validation, committing the agent generates its configuration artifacts (the agent config, metadata, deployment bundle definition, and — when enabled — the Streamlit app and Slack binding), commits them to the environment’s git branch, and dispatches the deployment.

In the create form, Create Agent performs the commit-and-deploy in one step. Deployment runs a workflow that registers the model and brings up the components:

Component When
Agent serving endpoint Always
Streamlit app When the Streamlit integration is on
Slack channel When the Slack Endpoint integration is on

Deployment cannot run before artifacts exist; if you try to deploy without generated artifacts, Nexa blocks it. Track progress in the Deployment step or the agent detail’s Deployment tab, either as a component Summary or the full Workflow Details logs.

Nexa agents move through three environments, each bound to a git branch:

Environment Branch
dev develop
test (also called staging) release
prod main

A running Nexa instance is bound to one environment. On the agent detail and deployment views, the environment determines which endpoint, app, and logs you’re looking at.

Once deployed, use the play/pause control on the agent card or row to manage its compute. It opens a modal listing the agent’s components and any dependent automations.

  • Stop is atomic — it stops the serving endpoint, Streamlit app, and Slack bindings together. Confirm with Stop All.
  • Start is selective — the agent endpoint always starts, and you choose which additional components (Streamlit app, Slack channel) to bring up. Confirm with Start Selected.

Start and stop poll for status, so you can close the modal and keep working while components come up or down.

Promotion moves a specific agent version to the next environment. Select Promote from the agent’s overflow menu.

  • Promotion is stepwise: dev → test, then test → prod. You can’t skip from dev to prod.
  • Promoting opens a pull request from the source branch to the target branch. The target team reviews and approves the PR in GitHub, which triggers the deployment into that environment.
  • Promoting to prod additionally requires the version to be deployed and healthy in test.

After you request a promotion, Nexa confirms it and opens the deployment progress view for the target environment. Promotion depends on your GitHub Integration and environment configuration.

Nexa deep-links to the platform-hosted logs and traces rather than streaming them itself:

  • Deployment logs / Workflow Details — the deployment workflow run, with per-job and per-step status and a link to the workflow run.
  • Traces — the agent’s traces, opened from the Traces action when available.
  • Component logs — the Streamlit app and Slack route link out to their hosted logs from the deployment progress panel.

Deleting removes an agent version from the current environment. The confirm modal lists what will be removed — bundled Streamlit apps, Slack channel bindings, and dependent automations (which cascade). Confirm with Confirm delete; the agent shows Pending_Delete, then Deleted once the cleanup workflow completes. Deletion is environment-local — it only affects the environment you’re bound to.

Agents deploy as Databricks model serving endpoints; the Deployment logs open the deployment workflow run and Traces open the endpoint’s traces in the Databricks workspace.