What today's agent news asks of a business choosing tools

Most of today's announcements circle the same subject from different sides: the software agent that acts on your behalf. The questions underneath them are older than the technology — who is accountable for the work, what it costs, and where a person still has to step in.

Agents have moved quickly from demonstration to deployment, and the conversation has moved with them. A month ago the interesting question was whether an agent could complete a task. Today the interesting questions are how much authority it should hold, what it changes about the people around it, and when it hands the work back. Those are not research questions. They are the questions a business has to answer before it lets an agent touch a customer, a ledger or a production system.

Microsoft puts AI agents in the insider-threat column

In security, an insider threat is not a villain. It is a trusted actor who already holds legitimate access and can still cause harm — through a mistake, through being manipulated, or through doing exactly what it was told in a situation nobody anticipated. The reason the term matters is that you cannot defend against an insider by keeping them out. They are already in. You defend instead with least privilege, with monitoring, and with a record of what was done.

Microsoft is now applying that same lens to AI agents. Its guidance is that enterprises should closely watch the agents that employees use with systems the AI labs do not control [1]. The framing in the headline is blunter still — that frontier models, in this specific sense, cannot simply be trusted. The point is not that agents are dangerous. It is that an agent with real access deserves the same treatment as any other trusted actor with real access: a narrow set of permissions, a log of its actions, and a person who can review them.

That is a design discipline more than a product. Give an agent the least access that lets it finish its job. Keep a record it cannot edit. Make sure a human can see what it did and undo it. We built 360REV so that every action an automated process takes is permissioned and recorded in the same way a person's action is, which is the point of the audit trail nobody thinks about. If you are introducing an agent anywhere near your data, the first thing to settle is not what it can do but what it is allowed to touch.

Anthropic's own workforce data shows where agents stall

The most useful corrective to agent hype today came from Anthropic's own numbers. A report drew on internal headcount data, agent session telemetry, calendar data, Slack patterns, and hiring scorecards, plus seventeen employee interviews [2]. That is a company studying what agents have actually done inside its own walls, not a vendor describing what they might do inside yours.

The headline figures the write-up pulls out are worth sitting with. Agents log roughly six times the working hours of humans. Junior technical hiring fell from 27% of hires to 9%. And agents spend about 80% of their time waiting on a human. That last number is the one a tool buyer should remember. Even inside an organisation built around this technology, the agent is not an independent worker. It is a fast, tireless assistant that stops and waits most of the time for a person to decide, approve or clarify.

That shape changes how you should plan. If an agent is idle four-fifths of the time pending a human, the bottleneck you are buying is not compute — it is the handoff. The question stops being "can the agent do this" and becomes "how smooth is the moment it hands back". A tool that makes the handoff clean is worth more than one that claims more autonomy and delivers more confusion. This is the same boundary we wrote about in what AI should and should not do in your business: the work an agent is good at and the work it should route to a person are different categories, and pretending otherwise costs you.

Personal agents may add contact-centre work before they remove it

There is a comfortable assumption that as customers get their own agents, your support load falls. Salesforce argues the opposite is the near-term reality. A personal agent can start a customer's request, but the business still has to finish it [3]. The company's framing is that personal agents will drive more contact-centre traffic than they offset.

The logic is straightforward once you say it plainly. A customer with an agent will initiate more contacts, more often, because the cost of starting one has dropped for them. But the hard part of most requests — the part that needs your data, your policies and your authority to act — still lands on your side. So the volume arriving at your door goes up, and the share of it that an agent on your side can resolve without a person does not rise fast enough to cancel that out.

For anyone choosing support tooling, this reframes the buying decision. The capacity you need is not just more automated answers. It is the ability to absorb more inbound of every kind and still route the genuinely hard ones to a person with full context. If a customer's agent and yours are both involved in one request, the record of that conversation has to be one record, not three.

What AI is doing to SaaS competition and pricing

Stripe surveyed 200 platform leaders and reports that AI is creating more SaaS competition and that platforms are feeling the pressure; the survey shows how they plan to evolve pricing, grow revenue, and retain customers [4]. The detail worth noting is the pairing of pricing with retention. When a category gets more crowded, the instinct is to compete on price. But a plan is a promise about what a customer gets, and cutting the price without changing the promise usually trains customers to leave when a cheaper promise appears.

If you sell software, the signal here is to watch your own pricing the way your buyers will — against a widening field. That is less about matching a competitor's number and more about being clear on what your plan actually delivers, which is a habit we have argued for before.

Natural-language analysis moves into the spreadsheet

Google is adding a quieter but practical capability: after connecting to BigQuery data with Connected Sheets, you can use Gemini to interact and analyse that data conversationally, right in Google Sheets [5]. The pattern matters beyond this one feature. Analysis is moving into the tool people already have open, and the barrier between "I have a question" and "I have a query" is thinning.

The convenience is real, and so is the trade-off. A conversational query is easy to ask and easy to misread. The system will answer whatever you asked, which is not always what you meant. The discipline that keeps this useful is the same one that has always separated a number you can act on from a number that merely looks confident — knowing what was counted, over what period, and with what excluded. That is the whole subject of numbers that change a decision, and a plain-language interface raises the stakes on it rather than lowering them, because it removes the friction that used to make people slow down.

The thread, pulled tight

Four of today's five items are arguments about agents, and they do not contradict each other. Microsoft says watch them like insiders. Anthropic's data says they wait on you most of the time anyway. Salesforce says they will raise your volume before they lower it. The common instruction is unglamorous and consistent: keep the human in the loop, keep the permissions narrow, keep one clean record of what happened. The business that treats agents as powerful but accountable colleagues — rather than as either magic or menace — is the one that will get the use out of them. For the rest of the week's announcements, our earlier daily briefing covers the ground before today.