This article presents a practical walkthrough for building a production-ready agent using Gemini Enterprise‘s Agent Platform and the Agents CLI, addressing the common problem of AI projects stuck in prototype mode due to context-switching between multiple tools. The tutorial demonstrates building an Industry Watch agent that reconciles semiconductor company press releases against SEC filings, covering the full lifecycle from setup to publishing.

The core idea is that the Agents CLI provides a set of skills that allow a coding agent to drive the entire agent lifecycle without leaving the coding environment. The walkthrough is structured in six stages: (1) Setup: install Agents CLI and Developer Knowledge MCP to give the coding agent platform-specific skills, with the note that the code-execution sandbox is only available in us-central1. (2) Build: describe the agent to the coding agent, which scaffolds an ADK project using agents-cli create industry-watch --agent adk --prototype and generates three deterministic tools — a press release fetcher, a disclosure fetcher hitting a real SEC endpoint, and a reconciliation tool that joins the two live sources without model inference, using CIK/ticker matching, date window, deduplication, and materiality scoring against the 8-K item taxonomy. (3) Deploy: use agents-cli deploy to host the agent on a managed runtime with automatic scaling, persistent sessions via Memory Bank, and a code-execution sandbox that keeps deterministic Python isolated from the model. (4) Govern: assign a scoped identity via Agent Identity, restrict egress to approved hosts through Agent Gateway, and protect the tool boundary with a Model Armor template that screens inputs and outputs for prompt injection and jailbreak attempts. (5) Evaluate: run automated pass/fail tests using a grounding metric that measures whether every claim is traceable to a real source, enabling CI integration. (6) Publish: deploy the agent to Gemini Enterprise via agents-cli publish gemini-enterprise, requiring an existing Gemini Enterprise app and a license, so that business users can query the agent in the same app they already use.

The article emphasizes that the architecture — not a bigger prompt — is what makes the agent reliable. The deterministic join between two live data sources is the product, and the model only narrates the result. The agent is built entirely through prompts to the coding agent, which executes the Agents CLI commands. The result is a governed, published enterprise asset that can be queried weekly with grounded, cited answers. Stated boundaries: the code-execution sandbox is us-central1 only; publishing requires an existing Gemini Enterprise app and a license. The tutorial assumes the developer has a coding assistant with Agents CLI installed. The example is specific to semiconductor stocks, but the pattern is generalizable to any task requiring live data, auditable answers, and a defended tool boundary.