Create Agent

Use Google's maintained agents-cli to create or enhance an Agent Development Kit (ADK) application.

Important Scope Check

This command creates a runnable Google agent application. It does not create a Claude Code subagent file under agents/.md or .claude/agents/.md.

• For a Claude Code subagent definition, use /agent-creator.
• For a Google ADK application, continue here.

Do not treat Agent Starter Pack as the default for a new project. Google placed that project in maintenance mode and moved active development to google/agents-cli.

Prerequisites

• Python 3.11 or newer
• uv and uvx
• Node.js, because agents-cli setup installs coding-agent skills
• A project name and whether this is a new or existing codebase
• Google Cloud only for deployment; local development can use an AI Studio key

Use command -v only to inspect whether the required executables exist. If a prerequisite is missing, show the official installation link and stop before changing the machine.

Instructions

1. Use AskUserQuestion to confirm whether the user wants a new ADK project or wants to enhance an existing one. Confirm the target directory before writing files.
2. Read the target repository's instructions and inspect its current files with Read, Glob, and Grep. Never overwrite an existing application blindly.
3. Inspect the installed CLI contract before constructing flags:

   uvx google-agents-cli --help
   uvx google-agents-cli create --help

1. Show the exact scaffold command and destination. Run it only after the user approves the write:

   uvx google-agents-cli create PROJECT_NAME

1. For an existing agent project, inspect the repository first, then use the current enhancement command shown by CLI help. Current releases expose:

   agents-cli scaffold enhance

1. Enter the generated project and use its own checked-in configuration as the source of truth. Do not invent template names, deployment flags, model IDs, regions, or environment variables.
2. Run the generated verification lane. The current CLI provides:

   agents-cli install
   agents-cli lint
   agents-cli run "smoke test"
   agents-cli eval run

1. If deployment is requested, authenticate explicitly and inspect agents-cli deploy --help. Present the target, Google Cloud project, region, identity, secrets plan, and cost implications before executing deployment. Supported deployment guidance currently covers Agent Runtime, Cloud Run, and GKE, but availability and flags may change by release.
2. If publishing to Gemini Enterprise is requested, inspect agents-cli publish gemini-enterprise --help and require a separate approval before the external mutation.

Output

Report:

• project name and absolute destination
• new-project or enhance mode
• exact agents-cli version and commands used
• files created or modified
• local lint, smoke, and evaluation results
• deployment or publication status, clearly distinguishing planned from run
• any remaining authentication, quota, or configuration work

Safety and Approval Boundaries

• Never write until the destination is confirmed.
• Never run agents-cli setup, install system dependencies, authenticate, provision cloud resources, deploy, publish, or initialize a remote repository without showing the exact action first.
• Never put API keys, service-account JSON, access tokens, or credential paths in generated source control. Use the generated secret-management pattern.
• Never use --allow-unauthenticated unless the user explicitly requests a public endpoint and acknowledges the exposure.
• Do not quote fixed prices. Link to current Google Cloud pricing for the chosen services because model and infrastructure prices change.

Error Handling

Resources

• Google agents-cli
• agents-cli documentation
• Google ADK documentation
• Agent Starter Pack migration guide