Google Cloud unveils Gemini Enterprise agent workflow

Google Cloud has outlined a workflow for building and managing AI agents through its Gemini Enterprise Agent Platform. It is designed to cover development, deployment, governance, testing and publication in a single process.

To show how the system works, Google used a semiconductor research tool called Industry Watch. The example agent compares company announcements and news coverage with filings made to the US Securities and Exchange Commission, using live data rather than a model's stored knowledge.

Many AI projects fail to move beyond the prototype stage because developers must switch between coding tools, cloud consoles, identity settings and deployment systems, Google said. Its answer is a command-line tool called Agents CLI, intended to let developers manage the full agent lifecycle from within a coding assistant.

The workflow starts by installing lifecycle skills and connecting what Google calls a Developer Knowledge MCP, which lets the coding agent refer to current platform documentation. According to Google, this is meant to reduce errors caused by models relying on outdated training data when working with fast-changing software platforms.

Agent design

In the Industry Watch example, the build process begins with a prompt asking a coding agent to scaffold a new ADK agent for a group of semiconductor stocks including NVDA, AMD, INTC, MU and AVGO. The generated project is then expanded with three deterministic tools: one to fetch SEC 8-K filings, one to gather public claims from news and investor relations feeds, and one to reconcile those two sets of records.

Google argued that this architecture is necessary for tasks where a chatbot alone would be unreliable. In this case, the system needs current information, direct links to source material and a way to join company claims with regulatory disclosures that the model cannot invent.

The reconciliation tool is central to that approach. It groups records into matched, filing-only and claim-only categories, removes near-duplicate news items and scores materiality based on 8-K item taxonomy, ranking more consequential disclosures ahead of routine announcements.

In this design, the model's role is limited to narrating the results returned by those tools. Google said this is intended to prevent the system from fabricating links between press releases and filings.

Managed deployment

Once built locally, the example agent is deployed to a managed runtime. Google said this turns a local script into a service that can support repeated use, retain context and isolate deterministic processing from the model.

The deployment stage also adds state management through AI Sessions and a Memory Bank. In the Industry Watch example, that allows the system to remember a user's watchlist, sector focus and preferred briefing format across sessions.

Google said deterministic Python tasks such as reconciliation, deduplication and materiality scoring can then be moved into a managed code execution sandbox. That keeps the analytical logic separate from the model while preserving the same user-facing behaviour.

Security controls

Governance features are a significant part of the workflow. Google said developers can assign a dedicated identity to each agent and restrict permissions to a narrow set of roles, rather than letting the agent inherit broad access rights.

Google also described network restrictions through Agent Registry and Agent Gateway, with outbound traffic limited to approved hosts. In the Industry Watch example, those sources include sec.gov, the GDELT project and investor relations feeds.

Another control comes from Model Armour, which screens prompts, model responses and tool output for prompt injection and jailbreak attempts. Google framed this as a safeguard against manipulated headlines or hostile source material that could alter an agent's instructions.

Testing and publication

Google also set out an evaluation process built around automated testing rather than informal inspection. In the example, the agent is tested on multi-turn analyst queries and graded on task success, tool use and hallucination.

One deterministic test checks that every accession number and 8-K item code cited by the agent appears verbatim in the underlying tool output. Google said this is intended to turn concerns about hallucination into a pass-or-fail condition.

Prompt changes should also be assessed against a baseline to ensure there is no regression in grounding before updates are kept, Google said. These evaluations can be integrated into continuous integration pipelines so weaker prompt changes are identified before release.

The final stage is publication in the Gemini Enterprise app, where business users can access the agent alongside other workplace tools. In Google's example, an analyst can ask which developments affected a semiconductor watchlist in the past week and which public claims lack support from an SEC filing. The result is returned as a cited answer that highlights the claim-only category.

Google said the same pattern could be applied to other use cases that depend on live data, auditable responses and strict boundaries between tools and models.