# What is a deployment agent?

Learn how a coding agent can inspect, plan, deploy, diagnose, and redeploy an app through a CLI with stable contracts.

## The short version

A deployment agent is a coding agent that can carry an application from local source to a live release using a deployment tool with machine-readable commands and results. With Ample, the agent can inspect the repo, resolve a plan, deploy services and data resources, read logs, and reuse the same workflow for later changes.

## Who this is for

This guide is for builders evaluating how AI coding agents can handle deployment work while keeping the source and infrastructure plan visible.

## How to do it

1. **Give the agent a scoped setup path.** The Ample homepage creates a one-use onboarding secret and a snippet that installs the CLI, authenticates it, and establishes Ample as the repo's deploy tool.

2. **Plan from repository evidence.** The agent runs Ample's repo inspection. Open questions include evidence, options, a safe default, and a named answer path.

```
   $ ample plan
   ```

3. **Record the infrastructure contract.** Writing ample.toml makes services, databases, volumes, links, domains, and non-secret configuration reviewable in the repo.

```
   $ ample plan --write
   ```

4. **Apply and diagnose.** The agent runs the synchronous deploy once, reads the final result, and uses build or runtime logs if the release fails.

```
   $ ample up
   ```

5. **Reuse the workflow.** On later changes, the agent uses the existing plan and content-addressed deploy behavior instead of inventing a new release process.

## What this supports

- Stable JSON output and exit codes
- Evidence-backed planning questions
- Version-matched agent documentation
- Scoped tokens and budget controls

## Know the limits

- The agent still acts with the permissions granted to its token
- Missing secret values require a human or external secret source
- Deployment automation does not make untested application code correct

## Common failure modes

- **The agent polls a synchronous deploy** Run the deploy once and read its live, failed, or blocked result.
- **The agent guesses through a plan question** Use the supplied evidence, options, safe default, and answer path.
- **The agent commits a secret** Declare the key as secret and pass its value with encrypted environment input.
