On October 8, 2026, Thomas Kurian stood on stage at Gemini at Work and described a piece of software that hires. Google Cloud's new Gemini agent, which the company calls "your new single, universal agent for work," can take an objective, break it into pieces, and create a roster of temporary, job-specific sub-agents to do the pieces. Each sub-agent gets its own identity. Some agents become persistent "coworker agents" with their own email addresses at an @agents.company.com domain, their own storage, their own calendar, and their own name in a document's version history. According to Google's own write-up of the keynote, the work can run for hours or days after you close your laptop.
Read the announcement once as a product launch. Then read it again as an HR document. A system that decides, mid-task, how many workers it needs, how senior they should be, what they may touch and when they stop existing is making staffing decisions. It makes them in milliseconds, thousands of times a day, and nobody in your finance or people functions signed any of them.
That is the shift in multi-agent systems for business operations this month. The orchestration layer now holds a power that every company on earth has spent a century fencing in with approval chains: the power to add labour. My argument is simple and a little uncomfortable. Your delegation of authority matrix, the boring spreadsheet that says who can sign a $50,000 purchase order and who can open a new role, has a hole in it the exact shape of an orchestrator agent. Fill the hole on purpose or your vendor's defaults will fill it for you.
The requisition nobody signed
Every company of any size has a ritual for creating a job. A manager writes a requisition. Finance checks it against the budget. Someone senior signs. HR sets the grade, the title, the access. IT provisions a laptop and an account. Months later, if the role ends, someone revokes the access, which is the step everyone forgets.
The ritual is slow on purpose. Adding a worker is one of the few decisions that compounds. The worker costs money every month, touches systems, makes mistakes that other people inherit, and leaves behind work product that someone must own. The friction in the requisition process is how the company protects itself from its own enthusiasm.
Now look at what Google described. The Gemini agent spins up sub-agents dynamically. Each one is, in Google's words, "cryptographically attested and governed like an employee, with least-privilege permissions." Security administrators approve the role-based permissions in advance. Every action lands in an audit trail attributed to the agent rather than to a person. Traffic flows through something Google calls Agent Gateway, an "AI network firewall" that enforces policy in real time.
That is a careful design, and the identity and access half of the requisition ritual has clearly been rebuilt for machines. The other half has gone missing. Nobody approved the headcount. The question "should this task have three workers or thirty?" got answered by the orchestrator, at runtime, based on its own read of the objective. The grade question, which model runs each sub-agent, got answered by a router. The lifespan question got answered by whenever the task finished, or whenever the money ran out.
Multi-agent systems for business operations moved hiring into the runtime
Google was not alone this month. The pattern showed up across every serious enterprise agent launch between late September and late October.
On September 29, 2026, OpenAI used DevDay to put the Codex harness behind a managed Agents API. InfoQ's recap reports that the API now incorporates "multi-agent capabilities from Codex," along with tool search, computer use and context compaction, with OpenAI running the execution infrastructure. The Codex CLI got an /agents interface for delegating and monitoring multiple tasks at once. OpenAI also launched Dots, persistent agents that keep working on ongoing tasks.
On October 5, 2026, Cohere shipped North 2. SiliconANGLE's coverage describes a rebuilt orchestration system that runs multistep agent tasks on its own and pulls in a person for steps that need sign-off. The admin console tracks token use down to the individual user and the individual agent, lets administrators set consumption tiers for users and groups, caps usage across the company and sends alerts before teams hit a limit. Cohere also added "autonomy policies" that limit each agent to its authorized actions.
On October 22, 2026, the Agentic AI Foundation announced that Anthropic had contributed Agent Skills and OpenAI had contributed Open Responses as open projects. Unite.AI reported that Agent Skills is supported by more than 70 agent platforms, with over 1 million skills published and nearly 280 million installations. Skills are the job descriptions of the agent world: packaged instructions, scripts and resources an agent loads when it needs them. When the job description becomes a portable file that 70 platforms can read, the cost of standing up a new specialist worker drops toward zero.
Put those four events side by side and the direction is plain. The vendors have built three things: identity for agents, permissions for agents, and spend meters for agents. What none of them can build for you is the rule for when an agent should exist at all.
The numbers are already headcount-sized
Google's keynote material listed customer agent counts that read like workforce reports. SOMPO has more than 10,000 custom agents across 34,000 employees. Orange Spain has more than 1,000. Tata Steel deployed more than 300 agents in nine months. The U.S. Chief Digital and AI Office reports more than 100,000 custom agents built on Gemini Enterprise. Nearly 500 Google Cloud customers each processed more than one trillion tokens in the past year.
Those are the persistent agents, the ones somebody built on purpose. The temporary sub-agents an orchestrator creates on the fly do not show up in those counts at all. They are born for a task and they die with it. If you asked most CFOs today how many workers their company employed last Tuesday, counting every sub-agent that existed for any part of the day, they could not tell you. Most could not tell you which system holds the answer.
Fan-out has a shape, and the shape decides the outcome
Here is the part that turns this from a governance nicety into a performance problem. Spawning more agents does not reliably make work better. Sometimes it makes work much worse, and the difference depends on the structure of the task.
In January 2026, researchers from Google Research, Google DeepMind and MIT published a controlled study of 180 agent configurations, "Towards a Science of Scaling Agent Systems". They held tools and token budgets constant and changed only the architecture: one single-agent baseline and four multi-agent designs, across three model families and four benchmarks. The Google Research blog summary reports results that every operating executive should tape to a wall.
On Finance-Agent, a task that splits cleanly into parallel pieces, centralized coordination "improved performance by 80.9% over a single agent." On PlanCraft, a task where each step depends on the one before it, "every multi-agent variant we tested degraded performance by 39-70%." The researchers also measured how errors spread. "Independent multi-agent systems amplify errors by up to 17.2x," while centralized systems "contained this amplification to just 4.4x." As tasks need more tools, they found, "the 'tax' of coordinating multiple agents increases disproportionately."
The same research group then built a model that picks the right coordination strategy for 87% of unseen task configurations. So the decision can be made well. It can also be made badly, and the penalty for making it badly on sequential work is severe.
Now return to the Gemini agent, the OpenAI Agents API and North 2. In all three, the orchestrator decides the fan-out. The vendor's orchestrator, tuned on the vendor's sense of the average customer, decides whether your month-end close becomes one careful agent walking through the reconciliations in order or six sub-agents racing in parallel and stitching their answers together. Month-end close is a sequential process. Each journal entry depends on the balances the previous one left behind. It is precisely the task shape the Google study says multi-agent designs damage.
Your org chart already encodes this lesson for humans. You do not hand a single reconciliation to six accountants and ask them to merge results. You give it to one person and a reviewer. That wisdom lives in team structures that took decades to settle. The orchestrator has none of it unless you write it down.
Model choice is a pay grade
The second staffing decision hidden in the runtime is seniority.
Google's announcement leans hard on routing. The Gemini agent runs each job on the best-fit model across Gemini models and Anthropic's Claude models, with other private and open models planned. Google writes that "the best model for the task is not always the largest one," and offers "smart routing" that sends each workload to the model with the best performance at the lowest cost. Constellation Research reports that enterprises can put routing on autopilot or limit use of expensive models. At DevDay, OpenAI pitched GPT-6.1 Sol as approaching GPT-6 Astra on several evaluations at one-fifth of Astra's token prices, according to InfoQ.
Translate that into workforce terms. The frontier reasoning model is the partner. The small fast model is the analyst. Routing is the staffing manager deciding who works the file. Every professional services firm knows that this decision sets both the margin and the risk on an engagement. Staff a complex matter with too many juniors and the client finds the error. Staff a routine matter with partners and the firm loses money.
When routing runs on autopilot, the vendor's cost function is now your staffing policy. The vendor optimizes for performance per dollar on whatever it can measure. Your exposure lives in what it cannot measure: the contract clause with a regulatory consequence, the customer email that sets a legal precedent, the forecast assumption a board will read. Those tasks may look simple to a router. They deserve the partner anyway, because the cost of a wrong answer dwarfs the token bill.
Google's own blog says per-token prices have dropped 98% since 2024. Falling prices make the grade question more important, because they remove the natural brake. When every model was expensive, nobody over-staffed. When the analyst model costs almost nothing, the orchestrator can afford to spawn twenty of them, and the Google study suggests twenty independent analysts can multiply each other's errors.
Spend caps are a hiring freeze applied after the money is spent
The vendors' answer to runaway fan-out is the cap. It is a reasonable first answer and a poor last one.
Google's caps live in the Cloud Billing Console. Gemini monitors token usage and sandbox costs, and when a project hits its cap, the project's agent pauses until someone resumes it with one click. Constellation reports that admins can set project-level caps now, with individual caps coming in the following weeks. Cohere's North Admin offers per-user and per-agent tracking, consumption tiers and early warnings.
Think about what a cap does in staffing terms. It is a hiring freeze, triggered by the payroll department, after the payroll has already gone out. It stops the bleeding. It tells you nothing about whether the spending was wise. It also stops the good work along with the bad: the project-level pause halts the agent mid-task, including whatever sequential process it was halfway through.
A one-click resume makes this worse in practice. Picture the person who sees the pause. They are usually the person waiting on the output, not the person who owns the budget. They click resume. The cap becomes a speed bump with a button on it.
Suppose a regional insurer sets a $5,000 monthly cap on a claims-triage project, and the orchestrator starts fanning each complex claim out to eight sub-agents instead of two because a model update changed its planning behaviour. The cap trips on the 19th. The claims manager resumes it, because claims are piling up. The overspend is now policy, and the root cause, a change in fan-out width that nobody approved, never appears on any report anyone reads. In this example the cap did its job perfectly and the company still learned nothing.
The useful control sits upstream of the meter. It governs the decision to spawn, and the meter only reports its consequence.
Writing the spawn policy
Every company running multi-agent systems in business operations needs a document that does not exist yet in most of them: a delegation of authority for agents. Call it a spawn policy. It answers, for each business process, the questions your requisition ritual answers for humans. Here is how I would build one.
Start from task shape, then grant width
Map your processes by structure before you map them by department. The Google study gives you the sorting key: parallel work benefits from coordination, sequential work suffers from it. Research synthesis, competitive scans, document review across a large set of independent contracts, and first-pass triage of a support queue are parallel. Financial close, regulatory filings, pricing changes, multi-step approvals and anything where step four consumes the output of step three are sequential.
Sequential processes get a single-agent lane, with fan-out disabled or capped at a width of one plus a reviewer. Parallel processes get a maximum width, written as a number, owned by a named person. The width is the headcount approval.
Limit depth separately from width
Width is how many sub-agents an orchestrator creates at once. Depth is whether those sub-agents can create sub-agents of their own. Depth is where the error amplification compounds and where the audit trail turns into a family tree nobody can read. For most operational work, a depth of one is plenty: the orchestrator may hire, and the hires may not. Any exception should require the same signature as opening a new team.
Set the grade ceiling and the grade floor
For each process, name the cheapest model allowed and the most expensive model allowed. The floor protects quality on work with legal, financial or customer consequences. The ceiling protects margin on routine work. Inside the band, let the router route. Outside it, the router should have to ask. This is how a law firm staffs: associates can do the first draft, a partner must sign the filing, and nobody bills partner rates for photocopying.
Give every worker an end date
Google's temporary sub-agents end with the task. Persistent coworker agents, with their own inboxes and Drive storage, do not. They accumulate context, permissions and data the same way a long-tenured employee does, without the annual review. Every persistent agent should carry a review date, a named human owner and an expiry that forces renewal. When the owner leaves the company, the agent's renewal should land on their manager's desk the same week, alongside the laptop return.
Count the workers you employed
Ask for a daily spawn log: how many agents existed per process, at what width and depth, on which models, for how long, at what cost. Cohere already tracks tokens per agent. Google already stamps a per-agent identity into every log and every virtual machine the agent starts. The raw data exists. What is missing is a person whose job is to read it the way a controller reads payroll, looking for drift. When the average fan-out on a process doubles in a week and nobody changed the policy, a model update or a prompt change did it, and you want to know which.
The org design question your vendor cannot answer
There is a reason every vendor stopped at identity, permissions and meters. Those are horizontal problems. Every customer needs them, and they can be built once and sold to everyone. The spawn policy is specific to your company. It depends on which of your processes are sequential, where your liability concentrates, which errors your customers forgive and which ones end the relationship. A vendor that tried to set these numbers for you would get them wrong for most customers, so the vendors ship sensible defaults and a settings page.
Defaults are decisions. When you accept autopilot routing and a project-level cap, you have adopted Google's or OpenAI's or Cohere's view of how many workers your processes need and how senior they should be. For many companies that view will be fine for most work. For the work that matters most, the sequential, high-consequence processes at the centre of finance, legal and operations, it will be wrong in exactly the way the Google study measured.
There is a second-order effect worth naming. Once spawning is cheap and automatic, the structure of your operations stops being something you designed and becomes something that emerges from thousands of runtime decisions. Your company will still have an org chart on the HR system. It will also have a second, invisible one, redrawn every few seconds by orchestrators, which determines how the work actually gets done. The companies that keep their performance as multi-agent systems scale will be the ones that made the invisible chart legible and put a human signature on its shape.
Say your company runs 40 operational processes through agents by the end of 2027. For illustration, if 15 of them are sequential and you leave fan-out on default, the Google figures suggest you could be running a third of your automated operations in the configuration most likely to degrade them, and you would see it only as a vague sense that the agents are "less reliable on the hard stuff." That vague sense is how most executives will experience the problem. It shows up in no single report. It is a structural choice nobody made.
What to do this quarter
The launches of the last 30 days compress the timeline. Google says the Gemini agent is available now, with industry editions in preview; Constellation reported general availability expected around late October or early November. OpenAI's Agents API is in public beta. North 2 shipped on October 5. Your teams will turn these on, because they work and because the demos are good. The spawn policy needs to exist before the first renewal cycle, while the configurations are still young enough to change.
Four moves, in order. Classify your top 20 automated or soon-to-be-automated processes as parallel or sequential, and put the result in writing. Assign a named owner to each, with authority over width, depth, grade band and agent lifespan. Turn on per-agent and per-process spend tracking everywhere the platform allows, and route the alerts to the owner rather than the user. Then hold a monthly review of the spawn log, run by finance, the same way you review headcount variance.
None of this requires a new tool. Every control I have described is either already in the products that shipped this month or is a policy document. What it requires is someone who understands both how orchestrators decide to fan out and how your business actually flows, and who can translate one into the other.
Architect the authority before the agents set it
Buying the Gemini agent, the OpenAI Agents API or North 2 gets you a capable workforce that hires itself. That purchase is the easy part, and your competitors are making the same one this quarter. The advantage goes to the company that decides, deliberately and in writing, which work may fan out, how wide, how deep, at what grade and for how long, and then checks every month that the runtime obeyed. That is an architecture decision about your operating model. It belongs to the CEO, the CFO and the COO, and it cannot be delegated to a settings page written for the average customer.
Agor AI Advisory does this work. We map your processes by task shape, write the spawn policy and the grade bands, wire the meters to the right owners and set up the review cadence that keeps the invisible org chart honest as models change underneath it. We have watched what happens when orchestrators staff work on vendor defaults, and we know where the 39% to 70% losses hide in a real operation.
Every week you run multi-agent systems without a spawn policy, software is hiring on your behalf under rules you never read. Take the authority back while the workforce is still small enough to count. Schedule a strategic consultation with us today.
Sources
- Google Cloud Blog, Gemini at Work 2026: Introducing Gemini agent, October 8, 2026
- Constellation Research, Google Cloud launches Gemini agent to work across enterprise systems, October 8, 2026
- SiliconANGLE, Cohere unveils North 2 AI agent platform with rebuilt orchestration and token spending caps, October 5, 2026
- InfoQ, OpenAI DevDay 2026 Recap for Developers, October 2, 2026
- Unite.AI, Agentic AI Foundation Adds Four Projects to Reach 10 in First Year, October 22, 2026
- Google Research Blog, Towards a science of scaling agent systems, January 28, 2026
- arXiv, Towards a Science of Scaling Agent Systems (2512.08296)