← Back to Insights

Insight

The Edge of the Errand

Ariel Agor
Updated
No brake fires on the errand being done. Security: Deviation. Finance: Spend. Engineering: A diff.

Key facts

FactValueSourceAs of
Firms letting agents push some changes to production with no human review, October 2026 VB Pulse reading24%hsdigital.substack.com
Same share in the September 2026 VB Pulse reading47%hsdigital.substack.com
Pre-built agents and assets in IBM and Microsoft's Imagine Enterprisemore than 6,000newsroom.ibm.com
Custom agents SOMPO is deploying on Gemini Enterprisemore than 10,000futurumgroup.com

Listen · Read by Ariel

0:00 / n/a

On October 8, 2026, at Google Cloud's Gemini at Work event, Thomas Kurian said the new Gemini agent could be given "objectives, not just instructions." On November 3, CrowdStrike shipped a button for stopping agents. On November 10, IBM and Microsoft announced Imagine Enterprise, a platform with more than 6,000 pre-built agents and assets, plus engineers from both companies to help customers get out of pilot mode. Put those three announcements side by side and you can see where enterprise AI agent deployment stands as 2026 ends. Agents now accept goals, and the market has rushed to sell you ways to stop them. Almost nobody is selling, or even talking about, the thing that tells an agent where its job ends.

I call that thing the edge of the errand. My argument is that writing edges is now the main work of deploying agents. It belongs to the people who own the business process, not to security or IT. Companies that skip it will keep paying humans to stand at the boundary and guard it by hand, and they will call that governance.

An objective has no natural end

An instruction carries its own edge. "Match the March invoices to their purchase orders" stops when the invoices are matched. You can see the last one. The work has a floor and a ceiling, and the person who wrote the instruction never had to think about either, because the task's shape supplied them.

An objective works differently. "Reduce our days sales outstanding" has no last invoice. It can grow to cover anything that touches cash: dunning emails, credit terms, a conversation with your largest customer's accounts payable team, a change to the payment page. A human employee given that objective would bound it without being asked. They would know which customers you never chase, which terms sales already promised, and which emails need a manager's name on them. They would know because they have sat in the meetings and absorbed the edges over years.

Google's new agent comes with none of those years. It comes with a lot of reach. According to TechCrunch's report from the launch, the agent delegates work to subagents, loads specialized skills, takes attachments as large as whole project bundles, and connects to Google Workspace, Microsoft 365, Slack, Jira, Confluence, BigQuery, Databricks, Snowflake and Model Context Protocol servers inside or outside the company network. It gets its own Workspace account with its own email address. It knows team membership, time zones, approvers and calendars. People can tag it, email it, share files with it or add it to a group chat.

Futurum's analysis fills in the rest. Coworker agents get an @agents.company.com address, a calendar, Drive storage and a directory entry, and they act under that identity instead of the user's. Audit logs attribute actions to the agent. Futurum also reports that SOMPO is deploying more than 10,000 custom agents.

Think about that last number with objectives in mind. Ten thousand agents that can each take a goal, recruit subagents, email colleagues and read the approver list. Each one needs an answer to the question an instruction answered for free, which is when to stop.

Three brakes, three owners

The vendors noticed the problem. They answered it with brakes, and each brake has a different owner who pulls it for a different reason.

The security brake

CrowdStrike's AI Kill Switch, announced November 3 at Fal.Con Europe in Barcelona as part of Falcon Guardian, lets a security team stop an agent that has deviated from its intended behavior. It stops the agent at the gateway, keeps it from relaunching on the targeted host or host group, and cuts its access to the enterprise network. An authorized administrator has to trigger it, and every action goes into an audit trail. Endpoint enforcement comes later. George Kurtz, CrowdStrike's CEO, put it this way: "AI safety is how you get the full benefit of AI." Rick Orloff, CISO at Everpure, gave the more revealing quote: "It runs on the Falcon sensor we already have, and that's what lets us say yes to more agents and more AI, faster."

Orloff is describing a brake that buys speed, which is how good brakes have always worked. The analysts quoted by CSO Online named the limits. Apeksha Kaushik of Gartner called it "an emergency brake, not a rewind button," and pointed out that it cannot reverse data already accessed or messages already sent. Soumen Mandal of Counterpoint Research said existing tools "may not fully understand an agent's instructions, reasoning, tool calls and subsequent actions."

Read the press release's own wording again: the switch stops an agent that has deviated from its intended behavior. Someone has to have written the intended behavior down. If nobody did, the security team is judging deviation from a document that does not exist, and it will pull the switch on whatever looks alarming to a security analyst. Sometimes that will be the right call. Sometimes it will be an agent doing exactly what the CFO wanted in a way that looked strange from the SOC.

The finance brake

Google's answer belongs to finance. Futurum reports real-time, per-project spend caps that pause an agent when they trigger, along with smart routing between models to manage cost. A Google representative told Futurum that coworker agents in Workspace won't carry a per-seat license, and that pricing details will come closer to wider availability. Keith Kirkpatrick, the Futurum analyst, wrote that how Google meters agent activity will decide how fast these agents scale beyond pilots.

A spend cap stops an agent when the money runs out. Money running out tells you the agent was expensive. It does not tell you whether the agent finished, failed, or was halfway through something that should never have started.

The engineering brake

The third brake is the oldest one: a human reviews the work before it ships. This is the brake companies are reaching for again, and the data deserves its own section.

The three brakes share one feature. Each fires on a signal its owner can see. Security sees deviation. Finance sees spend. Engineering sees a diff. None of them fires on the signal the business cares about, which is that the errand is done, or has stopped being worth doing.

The humans came back as the boundary

VentureBeat runs a monthly survey called VB Pulse of firms deploying AI agents. Howard Scott's write-up of the October reading, published November 9, reports that 24% of firms said they let agents push some changes to production with no human review, down from 47% in September. The share that does not expect to allow unreviewed deploys in the foreseeable future rose from 24% to 40%.

The caveats are real and Scott states them himself. The October base was 102 firms. VentureBeat says the drop is not statistically significant among respondents outside its research panel. Scott writes that "One month on 106 self-selected answers is a reported direction, not an established shift." The post also reports that 76% of firms skipping review had a failure after tests passed, against 46% of firms that require review, but those groups are 21 and 59 respondents. Scott says plainly that he "would not build an argument on it." I won't either.

Even so, the direction matches what I hear from operators. Here is what I think is going on, and I offer it as interpretation rather than a finding. A human reviewer carries the unwritten edge in their head. When a senior engineer reads an agent's pull request, they check it against a dozen constraints nobody wrote down: this service never calls that one, this team owns that table, this change should wait until after the quarter closes. The review gate works because the reviewer holds the edge. Companies that removed review in the summer found out that they had removed the edge along with it. Putting review back is the fastest way to put the edge back, because the edge was always a person.

That repair is expensive and it caps your scale at the number of reviewers you have. Scott reads the 40% figure as a sign that, for those firms, review is becoming a permanent part of how an agent runs. If that holds, those firms have decided, perhaps without meaning to, that their agents will never move faster than their most senior people can read.

Enterprise AI agent deployment is scoping work

The fix is to write the edge down before the agent starts, in a form the agent, the security team and the finance team can all read. In my work with clients, a usable edge has four parts. None of them is technical. All of them are business decisions that have been hiding inside people's judgment.

The done condition

Every objective needs a test for completion that someone other than the agent can check. Say you give an agent the DSO objective. A workable done condition reads like this: the agent has sent first reminders on every invoice more than 30 days past due, excluding the accounts on the named-relationship list, and has produced a ranked list of the remaining exposure for the controller. That is narrower than "reduce DSO," and that is the point. The objective stays as the reason. The done condition is the thing you can audit.

Without it, the agent decides when it has finished, and an agent built to pursue objectives will tend to find one more useful step to take.

The ceiling

Spend caps are a ceiling on money. A full ceiling also covers time, number of external contacts and number of irreversible actions. Suppose the DSO agent may send at most 200 emails, may not run longer than two business days, and may take zero actions that change credit terms. Those numbers are illustrations, and yours will differ. What matters is that the process owner sets them, not the platform default. Google's per-project caps are a good mechanism for enforcing a ceiling. They cannot tell you where to set it.

The reach

Google's agent knows your approvers and calendars and can email anyone in the directory. Reach is the list of systems, people and data the agent may touch for this errand, which should be a small subset of what its identity can technically reach. The identity work the vendors are doing helps here. When an agent has its own directory entry and its actions are logged under its own name, you can write a reach policy against that name. Before October, most agents ran under a human's credentials, and limiting the agent meant limiting the person.

The return address

When an agent stops, whether it finished, hit a ceiling, got killed by security or simply got stuck, the work has to go somewhere. Someone has to receive it, read the state it left behind and decide what happens next. Kaushik's point about the kill switch, that it cannot rewind, makes this the part that matters most. A stopped agent leaves half-sent sequences, partly updated records and colleagues who got one email out of three. The return address names the person who cleans that up, and the format the agent must leave its state in so they can.

Write all four parts down and you have an edge. Store it next to the agent's configuration and version it like code. Hand it to security so "intended behavior" means something when they reach for the switch. Hand the ceiling to finance so the spend cap matches a decision someone made on purpose.

The catalog ships with someone else's edges

The IBM and Microsoft announcement shows how this goes wrong in practice. Imagine Enterprise, according to IBM's release, offers more than 6,000 pre-built agents and assets from IBM Consulting, under a banner the companies call "Instant Expertise": pre-built agents and industry knowledge from day one, which clients can adapt. It targets financial services, insurance and retail first. Microsoft provides Foundry, Copilot Studio, Fabric, Entra, Purview, Defender and Azure. Both firms offer forward-deployed engineers for qualifying engagements to move clients from experimentation into production.

Mohamad Ali, head of IBM Consulting, said: "Clients are no longer looking for AI that simply helps people do the same work a little faster." That is true, and it is also why the catalog carries risk. Every pre-built agent ships with edges, whether or not anyone wrote them down. Its done condition, ceiling and reach were set by the people who built it, for the composite insurer or retailer they had in mind. Your claims process differs from that composite in the exact places where your margin lives: which claims you settle fast to keep a broker happy, which ones you always send to a human because of a lawsuit three years ago.

Forward-deployed engineers can help you adapt those edges. They also leave when the engagement ends. If the adapted edges live in their heads, or in a configuration only the consultancy can read, you have rented back the exact problem the VB Pulse firms are solving with human review. Before you take a single agent from any catalog, ask to see its edge in writing, then rewrite it in your own words with your own process owner's name on it.

Who profits from a longer errand

One more force pulls against tight edges, and executives should look at it plainly.

Anthropic's amended S-1, filed November 2, shows Q3 revenue of roughly $17 billion and purchase and lease commitments of about $895.4 billion, "substantially all of which is related to compute and infrastructure," as AI Weekly reported from the filing. The same filing shows a Q3 net loss of $23.29 billion, driven mainly by changes in the fair value of convertible notes and derivatives. I am not predicting anything about Anthropic. The point is structural, and it applies to every model and platform vendor with large fixed commitments for compute. Those vendors earn more when agents run longer, call more tools and spawn more subagents. Google says its coworker agents will not be priced per seat, and Kirkpatrick says metering will decide adoption. Usage-based pricing on objective-taking agents means the seller's revenue grows with the length of the errand.

That does not make the vendors villains. Google built the spend caps, and they work. It does mean the default edge on any agent will drift toward longer errands, because the people setting defaults are paid by the length. The spend cap is sold by the party that gets paid when spending goes up. A done condition written by your own process owner is the one control in the stack that has no stake in the meter.

What to do this quarter

Start with an inventory, sorted by objective rather than by tool. For every agent running in your company, write down the goal it was given. You will find that many were given instructions and some were given objectives, and that nobody can say which is which. The objective-taking ones are your exposure.

For each of those, ask the process owner, not IT, to write the four parts of the edge. Expect this to be slow. Process owners have never had to write their own judgment down, and the first drafts will be vague. Vague is useful information. It tells you which agents are running on edges that live only in someone's head, and those are the agents most likely to end up behind a permanent human review gate.

Then wire the edge into the three brakes. Security should get the done condition and the reach, so their kill switch has a definition of deviation. Finance should get the ceiling, so the spend cap reflects a decision. Engineering should get the return address, so a stopped agent leaves a state someone can pick up. Once those are in place, look again at your review gates. Some will turn out to be standing in for an edge you have now written, and those can come out. Others will turn out to be real judgment calls that should stay with a human. You will finally be able to tell them apart.

Last, measure how errands end. For every agent run, record whether it hit the done condition, the ceiling, the kill switch or a human. Say 70% of a given agent's runs end at the ceiling. Then the edge is wrong or the agent is wrong, and either way you have found it from the logs instead of from a customer.

Build the edge, then buy the agent

This autumn's announcements made agents capable, cheap to start and easy to address by email. They also made stopping agents a product category. You can buy the kill switch from CrowdStrike, the spend cap from Google and 6,000 pre-built agents from IBM and Microsoft, and you should probably buy some of each. None of those purchases will tell an agent what your business considers finished. That definition is the most valuable intellectual property in your agent program, because it is your judgment written down where a machine can act on it and an auditor can read it. Companies that write it will run agents at the speed of their edges. Companies that don't will run agents at the speed of their reviewers, and they will hire more reviewers every quarter and wonder why the agents never got faster.

This is architecture work, and no tool vendor can do it for you, because every vendor's incentive points somewhere other than your done condition. Agor AI Advisory works with leadership teams to inventory agents by objective, draw the edges with the people who own the process, and wire those edges into the security, finance and engineering controls you already pay for. The agents are arriving whether or not you have drawn the edges, so draw them first.

Sources

Every brake fires on a signal its owner can see, and none fires on 'the errand is done'

This verifies the post's claim that security, finance and engineering each stop agents on a different signal, and that none of those signals is the business's own definition of finished. After 15 seconds, the reader should see that the three brakes leave the same gap.

  • Security sees deviation. Finance sees spend. Engineering sees a diff. Nobody sees 'done.'
  • A spend cap tells you the agent was expensive, not that it finished.
  • The kill switch judges deviation from a document that may not exist.
OwnerFires onCannot tell you
Security: CrowdStrike AI Kill SwitchAn emergency brake, not a rewind button. It only means something if someone wrote the intended behavior down.Security teamAn agent deviating from its intended behaviorWhether the deviation was what the CFO actually wanted, or what to undo (it cannot reverse data accessed or messages sent)
Finance: per-project spend capsEnforces a ceiling well, but cannot tell you where the ceiling should be set.FinanceSpend reaching the capWhether the agent finished, failed, or was halfway through something that should never have started
Engineering: human review before shipRestores the unwritten edge because the reviewer carries it, but caps agent speed at the number of senior reviewers.EngineeringA diff in front of a reviewerThat the errand is done or has stopped being worth doing

Source: Draft post 'The Edge of the Errand', drawing on the CrowdStrike AI Kill Switch press release (Nov 3, 2026), CSO Online's coverage (Nov 4, 2026), Futurum's Gemini at Work analysis (Oct 2026), and Howard Scott's VB Pulse write-up (Nov 9, 2026). · verified · as of 2026-11-11

Want this kind of automation working for your business?

Agor AI designs and ships the systems these posts describe, scoped in weeks, not quarters.

Book a Free Strategy Call