Google's October 5 announcement consolidated its migration and modernization products into one portfolio built around agents, and one detail in it deserves more attention than the migration tooling itself: the agents stop.
The EKS-to-GKE Migration Agent, described in the Google Cloud Modernize announcement, handles discovery, translates Kubernetes manifests, and maps storage and networking across clouds. The announcement then lists, under a heading about enterprise-grade security, "Built-in Human-in-the-Loop (HITL) approval gates and in-memory credential security maintain strict GitOps compliance."
An agent did the mechanical work and a person kept the gate. Whatever your business migrates, that is the pattern to take from this announcement.
The portfolio pulls together infrastructure assessment, platform modernization and application modernization, with three named pieces. Modernization Hub is a new in-console experience where developers and architects "analyze source code, map dependencies, and accelerate modernization for Java, .NET, and mainframe applications". The EKS-to-GKE Migration Agent moves workloads out of Amazon EKS and is in public preview. The Agentic Quick Estimator turns VMware inventory exports into Compute Engine cost projections.
Notice the shape of each: the agent produces an artefact a person can check. A dependency map, a translated manifest, a cost estimate. None of them is a finished decision, and none of them commits the business on its own.
The industry conversation about agents has spent two years on capability and about six months on control. What Google shipped here is the third thing, and the one that actually changes a rollout: a division of labour where the agent absorbs the tedious, checkable work and the person spends their attention on the decision.
That division maps onto a distinction these posts have covered before. A site approval is not an action approval, as argued in browser agent approval versus action approval, and permissions alone do not tell you which actions need a person, as covered in Google ships agent controls. The migration portfolio is a worked example of the second idea: two mechanisms, one set of permissions and one set of gates, at the points where a mistake is expensive.
The rule underneath the table is reversibility. Agents run the stages you can redo, people hold the stages you cannot.
You will not run a Kubernetes migration, and the pattern still applies. Every business process that an agent is about to touch has cheap stages and expensive stages, and the mistake is treating the whole process as one block.
Take a quoting process. Gathering the site details, drafting the scope, pricing the line items and producing the PDF are all cheap and repeatable. Sending the quote to the customer, applying a discount beyond a threshold and committing a delivery date are not. An agent can own the first four and stop at the last three, which is the arrangement described in bounded permissions and human approval and the reason workflow automation for small business starts with a process map rather than a tool.
Two more pieces are needed for this to hold. Credentials must live in the platform rather than in the agent's context, which Google's announcement handles with in-memory credential security and Microsoft describes for Windows as containment plus identity so that organisations can "control what agents can access, understand which agent performed an action, and govern agents at scale" in its hybrid intelligence announcement. And the approval has to be specific, per OpenAI's guidance for agents that operate software, which keeps a person in the loop for purchases, authenticated flows and destructive actions and warns that content an agent reads cannot authorise an action.
For a process with no irreversible step, gates are friction. If the worst outcome is a bad draft that a person rewrites before anyone sees it, adding approvals buys nothing and slows the work. The split earns its place when the process touches money, customers, credentials or systems of record, and not before. That is also why secure deployment is the real barrier rather than budget, and why the unglamorous part of these projects is deciding which lane each step belongs in.
It makes the reversible stages much faster and holds the irreversible ones at human speed. If your process is mostly irreversible steps, an agent adds little, and that is useful to learn before buying anything.
A notification after the fact is not a gate. Last year's incident pattern is an agent that acted and told you afterwards. The decision has to sit before the action.
Fewer than you think, on the steps that genuinely cannot be undone. Gates that fire on routine work get clicked through within a week, which is the failure mode described in agent governance controls.
No. The migration announcement is a worked example of the pattern, not a prerequisite. A quoting process, an invoice run or a support desk has the same cheap stages and expensive stages.
Maxpertise builds and runs the AI team that operates your business, in a managed lane where we host and run the team, or an embedded lane where AI engineers work inside your existing team. Both lanes keep the two human gates: a person approves the plan before the work starts, and a person reviews the result before it ships.
The stage-by-stage split is the first thing we write down, because it decides what the agent may own and where your attention goes. It sits alongside the process map behind how to automate a back office without an IT team, and it is scoped the way spec-driven development with AI agents describes. If you want the method, start at how it works; the engagement options are on pricing, and contact with one process and its irreversible steps is enough to scope it.
[1] Google Cloud, Google Cloud Modernize: Accelerate transformation with AI, October 5, 2026. Read directly: Modernization Hub as "a newly launched in-console experience where developers and architects can analyze source code, map dependencies, and accelerate modernization for Java, .NET, and mainframe applications"; the EKS-to-GKE Migration Agent handling discovery, manifest translation and storage and network mappings; and "Built-in Human-in-the-Loop (HITL) approval gates and in-memory credential security maintain strict GitOps compliance."
[2] Google Cloud, Welcome to Gemini at Work 2026: Introducing the Gemini agent, October 8, 2026. Read directly on October 10, 2026 for the governance and cost controls.
[3] Microsoft, Building Windows for Hybrid Intelligence, October 7, 2026. Read directly: local and cloud AI combined for agentic experiences, Microsoft Execution Containers generally available on Windows, and the aim that organisations "control what agents can access, understand which agent performed an action, and govern agents at scale".
[4] OpenAI, Computer use integration: handle user confirmation and consent, accessed October 12, 2026. Human in the loop for purchases, authenticated flows and destructive actions; untrusted content cannot grant permission.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.