Let me first talk about something from a project.
An injection molding customer trialed an expediting agent in the first half of the year: it could read purchase orders and supplier delivery dates, identify which order would be late, and then automatically send emails, make phone calls, and adjust delivery dates when necessary. Two weeks after launch, the results were good, and the buyer said it saved two hours a day. In the third week, the warehouse overflowed.
The reason is not complicated. The agent only looks at delivery dates, not loading docks. This factory has only two loading docks and can unload at most twelve trucks a day. The agent moved all the delivery dates on eighteen purchase orders earlier, trucks lined up at the gate, the materials that could not be unloaded piled up in the temporary area, and once the temporary area was full, it blocked the fire access route. Throughout the whole matter there was not a single technical alert: permissions were correct, the model was correct, the number of calls was normal, and every operation was logged.
Later, I encountered the same shape in several projects: technical monitoring all green, while things went wrong at the business front line. So when I saw this wave of platform updates in November, what I cared about was not how much they can do, but which layer they install the switches at.
1. First, let's clearly lay out the three actions for this week
On November 16, UiPath updated its platform: coding agents moved to general availability and integrated with Claude Code, Codex, Cursor, and Google Antigravity; it launched the Delegate agent, which retrieves data across systems, modifies records, and attaches supporting evidence, while leaving important judgments and final approvals to humans; Data Fabric lets agents directly use data in SAP or Salesforce without copying it, while inheriting existing permissions; it added a runtime checker that checks in real time whether agents comply with enterprise rules while they work; Model Center can set permissions by model or tool; and the entire platform supports on-premises deployment on Linux. The official positioning in one sentence is: the same set of security policies applies to agents, automations, and employees at the same time.
On November 17, Microsoft Ignite opened in San Francisco. Before the event, the official documentation had already changed Agent 365's positioning: from an “agent control plane” to an “observability platform for AI,” with its four pillars being management, governance, security, and business value. On the governance side, it manages release approvals, owners, and connectable tools, with activity records going into Purview for audit review; on the business value side, it breaks down Copilot Credits consumption across four dimensions—user, group, agent, and service—and can also set spending caps. The documentation states explicitly: Agent 365 is licensed per person, not per agent.
The third piece of material is not news; it is data. Deployed but Ungoverned, released by ServiceNow and Phronesis Partners in October, surveyed 3,766 security and technology decision-makers at the vice president level and above across 20 countries and 8 industries:
| Number | Meaning |
|---|---|
| 90% | There has been an agent-related security incident in the past 12 months |
| 36% | Can bring these incidents under control within 24 hours |
| 34% | Automatically register new agents during deployment; the rest relies on manual entry. |
| 19% | Can see over 90% of agents in real time |
| 20% | There is a named individual who is fully responsible for agent governance |
| 56% vs 46% | Proportion self-rating 'safety has been established', versus the proportion with actual control. |
| 82% / 76% / 32% | Running agents in OT or physical environments / have had incidents there / have a separate governance framework |
Put these three things together, and they point to a rather specific gap: the platform has delivered the "can it be turned off" part, but "when it should be turned off" is still up in the air.
Numbers must come with qualifiers:The figures are self-reported by respondents, not independently measured; the publisher itself sells a governance platform; and subjective estimates like “how many agents can you actually see” tend to overestimate. So 90% and 36% work as trend signals, not as precise ratios to be written into a project proposal. What is genuinely solid are those two proportional relationships: more say they “have records” than say they “auto-register,” and more say they “have an access map” than say they “can see in real time” — the gaps in both cases are filled in manually.
II. The switches provided by the platform govern another kind of matter
Lay out the new capabilities built on both sides and look at them, and you'll find that what they actually govern is the "form of invocation," not the "business outcome."
| Platform capabilities | What is it looking at? | What can't it see? |
|---|---|---|
| Runtime Checker (UiPath) | Were any safety rules set by the company violated during the execution? | Is this operation reasonable from a business perspective? |
| Model Center / Tool Permissions (UiPath) | Which model was used, and which tools were allowed to be called | This order's amount, delivery date, and customer impact after the call |
| Unified tool toggle (Agent 365) | Whether MCP servers, plugins, skills, and connectors are permitted | Whether what it does after being allowed through is right or not |
| Agent Registry and Map (Agent 365) | What agents there are, who they belong to, and what they are connected to | How big a business loophole do these connections represent? |
| Subsidies and Quotas (Copilot Credits) | How much was consumed, who is using it, and whether it went over budget | What did all this spent money get in return? |
This is not a flaw of the platform; it is a division of labor. The platform provides general capabilities: allowing a tool to be permitted or prohibited, and allowing a call to be recorded, metered, and interrupted. It does not know how many trucks your company can unload at one dock in a day, nor does it know how much compensation is owed if a customer's order is delayed by a day. Only the business system knows these things.
So the answer to the question “where should the gate be installed” is clear: The switch is installed on the platform, and the criteria come from the business system. The line that triggers the interruption must be drawn on fields such as amount, quantity, capacity, credit, and delivery date.
III. Why Technical Rules Cannot Stop Business Incidents
If you break down the previous example of the injection molding factory, you will see three levels of misalignment.
The first layer is a misalignment of criteria. Technical rules look at “whether this tool is allowed to be invoked” and “whether this model is authorized” — all pass. Business rules look at “how many trucks can be unloaded per day at most,” but this constraint does not exist at all in the fields the agent can see. It does not know that its actions will hit a physical upper limit.
The second layer is a frequency misalignment.A person changing ten orders a day will weigh it themselves; an agent changing two hundred orders a day, as long as the rules allow it, won't feel anything is wrong. The same operation, once quantity changes to a certain degree, becomes an accident. This kind of thing cannot be stopped by approval—no one is willing to click approve two hundred times a day.
The third layer is a timing mismatch. Post-event audits can catch it, but by the time they do, the goods are already queued at the door. That is exactly what ServiceNow's 36% figure refers to: it is not that no one discovers it, but that it is discovered too late. What is truly useful is blocking it beforehand, logging an entry in-process, and being able to replay it afterward.
A counterintuitive lesson: the more "constraints" are written into business fields, the better the agent actually works. Because it knows where the boundaries are, the suggestions it gives are closer to a version that can be implemented; when the boundaries are unclear, it can only give a plan that looks reasonable, and then people have to discover that it doesn't work.
IV. Five Gates You Can Install in Odoo Right Now
The following five things do not require waiting for a platform upgrade, nor do they require purchasing new products; they can be implemented with standard features plus a small amount of configuration. The key is that each one must have clear criteria and a way to disable it.
| Gate | How to install | What criterion is used | What stops immediately after deactivation? |
|---|---|---|---|
| Dedicated account | Give each agent an independent user, and grant only the minimum permission group | the scope of business it is responsible for | Deactivate account; all actions immediately become invalid |
| Tool whitelist | Expose only the necessary methods (service actions or interface endpoints); leave the rest unexposed | The tasks it needs to complete | Calls not in the whitelist will directly throw an error |
| Recording rules | Use record-level rules to restrict it to only viewing/editing data within its own scope | Company, Department, Customer Ownership | Out-of-bounds queries return empty and cannot be written |
| Threshold circuit breaking | Add conditional logic to automation rules; route to manual approval when the threshold is exceeded | Amount, Quantity, Credit Balance, Available Inventory | Documents exceeding the threshold remain pending approval and are not saved to the database |
| Master Switch and Logs | System parameters serve as the master switch; enable change tracking for key fields | Launch Phase and On-Call Arrangements | Turn off the main switch, and all automatic actions stop |
The third "record rule" is the easiest one to bypass and deserves a separate reminder. In Odoo there is a commonly used construct called sudo(); its purpose is to skip permission checks when reading and writing data. Developers add it in server-side code to save effort, and once it is added, account permissions and record rules no longer take effect. Before connecting the agent, search the code for sudo() and confirm one by one whether each use is necessary—this step takes ten minutes and can save a lot of explanations later.
The fourth one, "threshold circuit breaking," only makes sense when applied to specific fields. Several commonly used ones: order amount (amount_total), customer credit limit and used credit (credit_limit and accounts receivable balance), available inventory (qty_available minus reserved), supplier historical on-time rate, number of changes per day. Write the conditions clearly and the actions clearly: when exceeded, is it suspended, downgraded to read-only, or requiring secondary confirmation—the three consequences differ greatly.
An actionable acceptance test: disable the account used by the agent, then have someone click the three operations it normally performs. If all three fail, the gate is real; if one still works, it has borrowed someone else's permissions or bypassed it in code. This test takes five minutes and is more useful than reading ten pages of architecture diagrams.
5. The OT side needs its own separate set
The most striking set of numbers in that ServiceNow group is this: 82% of organizations run agents in OT or physical environments, or have agents interact with them; 76% have had incidents involving agents in that environment; only 32% have established a separate governance framework for OT.
The reason is not hard to understand. IT governance logic is "permissions + logs + rollback," and none of these three fully holds in OT: a stopped production line cannot be rolled back, changing a process parameter has no undo key, and the consequences of equipment actions are physical. Therefore, the gate on the OT side cannot simply copy IT's approach.
People doing ERP must hold two points on this boundary. First, distinguish who writes what: changing a production order in ERP, changing the schedule of a work order in MES, and directly issuing parameters to equipment are three matters of different magnitudes, and cannot all be handled by one agent. Second, design 'degradation' in: the agent can suggest and draft, but the step that actually reaches the equipment must either go through manual confirmation or an independent control loop; it cannot rely on the same model to both propose the plan and execute it.
This point is especially important for projects integrating Odoo with MES. Odoo's manufacturing module manages orders, BOMs, operations, and work hours; the next layer down is MES's work orders and production reporting, and the layer below that is equipment. Where to draw the agent's permission boundary is best defined clearly at the blueprint stage, rather than relying on people to watch it after go-live.
Six, four questions, answers can be elicited in one hour
There is a very useful self-test in that ServiceNow survey: pick an agent that is already running in production, answer four questions within an hour, and the answers must come from the system, not from memory.
| Problem | Roughly where in Odoo should I check? | What does it mean if you can't answer? |
|---|---|---|
| Who approved it going live? | Approval Records, Discussion Records, Change Orders | There is no launch process, only verbal consent |
| What can it touch now? | Permission groups the account belongs to, record rules, tool whitelist | Permissions are accumulated; no one can explain them all. |
| What has changed about it since launch? | Field change audit trail and automation rule modification records | Changed but not recorded; if something goes wrong, there is no way to replay it |
| Who is responsible for it now? | Account Owner, Position Handover Record | The person is gone, but the account is still running. |
Among the four issues, the fourth is the easiest to overlook and the most likely to cause trouble. After staff turnover, an agent's account is still registered under a departed employee's name, the password sits in some shared document, and no one dares to disable it—because no one knows what disabling it would affect. Every company has a few accounts like this, but no one has ever counted them.
VII. Wrapping Up with Three Sentences
First, define the criteria before discussing capability. The question "What can it do?" has no answer unless we first answer "Under what circumstances is it not allowed to act?" The criteria must be written into business fields: amount, quantity, capacity, credit, delivery date. For scenarios where no criteria can be written, it means it has not been thought through whether automation is needed.
Second, the gate must be able to stop with one click, and someone must dare to press it. Technically, being unable to stop is an accident; organizationally, not daring to stop is also an accident. Assigning each agent a person who can make the final decision to disable it is more effective than installing three more monitors.
Third, don't manage things that go live every day with the cadence of quarterly reviews. The suggestion from that ServiceNow executive is a short daily meeting to look at deployments, risks, and anomalies. It sounds heavy, but if your agents change two hundred tickets a day, quarterly reviews really can't keep up.
Back at that injection molding factory. What was added later was not sophisticated at all: they added a “dock capacity” check to the agent—if the number of order changes in a day exceeded a certain number, it would be routed to a human; “changing the delivery date” and “changing the quantity” were split into two permissions, with the former delegated to the agent and the latter requiring human approval. After the changes, the purchasing clerk still saved two hours a day, and the warehouse never got congested again.
The platform will keep improving the toggle's usability; that's not our concern. What we need to worry about is something else: which field the line that should be triggered is drawn on.
Sources: UiPath 2026-11-16 platform update (reported by ZDNet Korea and others; official pricing and launch dates were not disclosed); Microsoft Learn Overview of Microsoft Agent 365 and the official Microsoft Ignite 2026 pages and agenda (2026-11-17 to 11-20, San Francisco); ServiceNow and Phronesis Partners Deployed but Ungoverned: The State of Agentic Security (2026-10; sample of 3,766 people, 20 countries, and 8 industries) and cited reporting by BizTech Magazine in November 2026. The enterprise scenarios in the text are examples that synthesize multiple projects, not a single customer; Odoo implementation methods are based on versions 19/20.
Zhongdi Information · Enterprise Management Digitalization Observation · Survey data are self-reported by enterprises; please refer to the original text when citing.
