Principal Basil C. Puglisi reads this week’s agent launches as a deployment question that sits one level above the technology: who decides when the agent is about to act. On October 8, Google Cloud CEO Thomas Kurian introduced the Gemini agent, a single, universal agent for work, at Gemini at Work 2026. Google says it plans the work, uses skills and tools, connects to a company’s systems, and brings back something finished. The same morning, Zone & Co introduced AgentHQ, which lets finance agents take approved write actions directly in NetSuite.
Both vendors put governance near the top of the pitch, and that’s progress. Google lists four questions every agent program has to answer, and Zone says finance teams must control what each agent can do, when it runs, and what requires human approval. Puglisi’s consulting position is that those controls describe the machinery well, while the deployment still has to name the person who holds authority at each decision point. A permission setting tells you what an agent is able to do. It doesn’t tell you who signed off on letting it.
That gap is the subject of Checkpoint-Based Governance, the framework Puglisi uses to put a named human at binding decision points over AI output, with a record of who decided and why. His September 29 essay, Human-AI Collaboration Needs More Than a Human in the Loop, made the same argument for everyday teams. A person sitting somewhere in the workflow is presence, and presence isn’t authority. This week’s launches make the point practical, because the agents now come with the hooks a checkpoint needs.
What Google and Zone announced on October 8
Google describes the Gemini agent as one interface and one API that answers questions, works autonomously on assigned objectives, and writes and runs code. It can be scheduled or set to respond to events, and it can create temporary sub-agents or act as a coworker agent with its own identity, email address, and storage. Google says it connects to tools such as Salesforce, ServiceNow, Jira, BigQuery, and Snowflake, and to any Model Context Protocol server.
On governance, Google frames the program around four questions: who the agent is, what it’s allowed to do, what it did and where that can be seen, and what it should never touch. Its answers are agent identities with least-privilege permissions approved by security administrators, an audit trail that attributes every action to the agent rather than a person, and an Agent Gateway that enforces written policy on traffic in, out, and between agents. Google also announced real-time spend caps that pause a project’s agent when a limit is hit, with work resuming by a single click in the console.
Zone’s AgentHQ launches with more than 10 specialist agents across bill matching, bill payment, and period close. The company says finance teams can configure, monitor, and stop agents, approve actions, and audit what has been done. “As agents take on more responsibility, containment, monitoring and control have to become part of the infrastructure, not something added after the fact,” said Billie Miric, Zone’s Global Head of Product.
Controls answer what an agent can do, checkpoints answer who decides
Google’s four questions are the right inventory, and Puglisi would hand them to any client as a starting checklist. Each one is answered with a system feature, though. Identity, permissions, audit logs, and a gateway policy can all be configured perfectly while nobody in the business owns the call to let an agent release a payment batch or send a contract. That’s where deployments drift. The agent is technically governed, and the decision is still orphaned.
The spend cap shows the pattern in miniature. When a cap triggers, the agent pauses and someone can resume it with one click. Checkpoint-Based Governance asks who that someone is, what they check before clicking, and where the reason gets written down. Zone’s approval step raises the same questions for money movement. An approval queue only works as a checkpoint when a named finance owner reviews against a stated rule and the record shows the decision and its basis.
Puglisi also flags the coworker agent model as the place where authority gets blurry fastest. Google says a coworker agent acts under its own identity and sees only what it’s given. That protects data. It also means an action can come from an agent that a whole team works with and no single person feels responsible for, unless the deployment assigns one.
How Puglisi would deploy a first production agent
The brief he recommends starts with a short decision map, written before any agent gets write access. For each workflow, list the actions that change something outside the agent’s sandbox: paying, posting, sending, signing, deleting. Mark each one as automatic, checkpointed, or forbidden. Forbidden actions go into the gateway policy. Checkpointed actions get a named owner, a rule the owner checks against, and a place where the decision is logged next to the agent’s own audit trail.
The second step ties the checkpoint to a measure. In Factics terms, each checkpoint carries a KPI defined up front, such as the share of agent proposals approved without edits, the time a proposal waits for review, and the error rate found after approval. Those numbers tell the team when a checkpoint can be loosened and when the agent needs tighter scope. A finance team piloting AgentHQ, for example, could start with every payment batch checkpointed and move the low-risk bill-matching corrections to automatic only after a stated run of clean approvals.
The third step is the stop. Zone says finance teams can stop agents, and Google’s spend cap pauses them. Puglisi would write down who can trigger a stop, who restarts the work, and what review happens in between, so the pause becomes part of the record.
What to watch as universal agents reach production
The vendors are converging on the same control stack: agent identity, scoped permissions, audit trails, policy gateways, and approval steps. That makes the remaining work organizational rather than technical. Puglisi would watch two signals over the next quarter. The first is whether customer case studies start naming the human role that approves agent actions, rather than only the hours saved. The second is whether platforms let an approval record carry the reviewer’s reason alongside the agent’s log. Until both show up, his guidance is simple: give the agent a scope, give each consequential action an owner, and write the checkpoint before the agent gets the keys.
Sources
- Google. (2026, October 8). Google Cloud launches Gemini agent [Announcement]. The Keyword. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-at-work/
- Kurian, T. (2026, October 8). Gemini at Work 2026: Introducing Gemini agent. Google Cloud Blog. https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026
- Zone & Co. (2026, October 8). Zone & Co introduces AgentHQ, bringing action and control to AI agents in the ERP [Press release]. PR Newswire. https://www.prnewswire.com/news-releases/zone–co-introduces-agenthq-bringing-action-and-control-to-ai-agents-in-the-erp-302901951.html
Questions readers ask
What is the Google Cloud Gemini agent announced on October 8, 2026?
Google Cloud describes the Gemini agent as a single, universal agent for work, introduced at Gemini at Work 2026. It answers questions, works autonomously on assigned objectives, and writes and runs code from one interface and one API, connecting to business systems such as Salesforce, ServiceNow, and BigQuery and to Model Context Protocol servers.
How does Google say it governs Gemini agents?
Google frames agent governance around four questions: the agent’s identity, its permissions, what it did, and what it should never touch. Its answers include attested agent identities with least-privilege permissions, an audit trail that attributes actions to the agent, an Agent Gateway that enforces written policy, and real-time spend caps that pause a project’s agent.
What does Zone AgentHQ do in NetSuite?
Zone & Co says AgentHQ launches with more than 10 specialist finance agents for bill matching, bill payment, and period close that can take approved write actions in NetSuite. Finance teams can configure, schedule, monitor, approve, stop, and audit agent work from one place.
What is Checkpoint-Based Governance?
Checkpoint-Based Governance is Basil C. Puglisi’s framework for placing a named human at binding decision points over AI output, with a record of who decided and why. It separates a person’s presence in a workflow from actual authority over the decision.
How should a business start deploying an AI agent with write access?
Start with a decision map that lists every action the agent can take outside its sandbox and marks each as automatic, checkpointed, or forbidden. Give each checkpoint a named owner, a rule to check against, a logged decision, and a KPI such as approval-without-edit rate, so the team knows when scope can safely widen.
#AIg
Leave a Reply