Enterprise agents converge on a single front door

An IT lead at a manufacturing group built a spreadsheet this year and sorted every AI tool in use by who bought it. By row twenty the list was still growing: the chat assistant inside the office suite, the coding tool the engineers picked, the service agent embedded in the CRM, the process agent finance built for itself. Each one had its own login, its own permission settings, its own monthly bill. When the fourth-quarter budget review came around, nobody could answer a basic question: across twenty tools, what did the company spend on AI this year, and what came back.

That question produces an obvious reaction inside a large company, which is consolidation. Rather than let twenty tools each keep their own memory and permissions, fold them into one front door with one identity, one audit trail, one invoice. At Gemini at Work 2026 on 8 October, Google put that route on the record: not another assistant, but a single agent that handles conversation, knowledge work, image generation, and writing and running code (Google Cloud announcement).

The wording of the launch gives away what it is trying to change

Google describes the agent as something you can chat with, assign an objective to, schedule, or trigger from an event. It runs on the web, on phones, on desktops and in a terminal, and it reaches into Workspace, Microsoft 365 and Slack, and into third-party apps as a headless service. In other words, it stops being a product you open and becomes a work object that is always on.

The dividing line sits in one phrase: you give it objectives, not instructions. An instruction is a step. An objective is a delegated outcome. Delegating an outcome forces three things on the system. It has to hold context on its own, it has to know which systems it is allowed to act on, and it has to keep working across days. Google’s answer is persistent execution in the cloud with one set of memories, one context and one personalisation graph, so a job that takes hours or days keeps running after the laptop closes and is where you left it when you come back.

Three pieces were pushed down into the product to support that promise. Multi-agent orchestration lets the agent spin up a temporary roster of sub-agents, each with its own identity, and coordinate parallel and sequential steps over hours. Coworker agents are the persistent version: a defined role that survives across days and sessions, with its own @agents.company.com email address, its own storage, and access only to the context a team grants it. Memory comes in four kinds: session, semantic, procedural and episodic. Around those sit a tools registry and a skills registry, so a department can publish the way it does a job as a reusable entry the whole company can call.

Making the model swappable tells you the vendor expects it to lose value

One sentence in the same post deserves to be pulled out on its own: Gemini is the agent, and the model underneath is a separate choice. It picks a model per job, orchestrating across the Gemini family and Anthropic’s Claude today, with other models later. The stated reason is blunt. The leading model changes every few months, so keeping the choice open means your context, your skills and your data stay put when it does (same announcement).

That is the most honest line in the release. A vendor that decouples the model is conceding that the model layer will keep commoditising and cannot hold a premium. Which leaves a short list of things that do hold it: context and memory, skills and tool definitions, and the records of what was executed and approved. The location of switching cost moves accordingly. In 2024 the worry was that prompts were tied to one model. In 2026 the real question is whose database holds the organisation’s memory, and whether it can be moved.

The benefit of consolidation is real, and it comes from removing surfaces

The problem with twenty scattered assistants is not that they are hard to use. It is that they cannot be governed. Twenty tools mean permissions spread across twenty admin consoles, audits that have to be chased vendor by vendor, and cost allocation that defaults to seat counts, which have nothing to do with who is actually using AI to get work done. Consolidating into one front door puts identity, permissions, audit and spend caps on one layer, which is the precondition for managing AI cost at all.

The governance features in this launch point the same way: cryptographically attested agent identity, OAuth-propagated permissions, action-level audit trails attributed to the agent, an agent sandbox and an agent gateway, plus real-time per-project spend caps that pause an agent when triggered (Futurum’s on-site analysis). The same architecture appeared in April, when Gemini Enterprise Agent Platform replaced Vertex AI and the new parts were Agent Identity, Agent Registry, Agent Gateway and execution traces (April launch). Registering identities, intercepting at a gateway, retaining traces: these are not assistant features. They are infrastructure features.

The scale numbers need a label. Google said nearly 90 percent of the Fortune 100 use Gemini Enterprise, and nearly 500 cloud customers have each processed more than a trillion tokens (Futurum). Both are the vendor’s own figures with no independent check, and using a product is not the same as running a critical process on it. What actually explains the pressure to consolidate is the buying logic behind those customer lists, not the percentages.

The costs sit in three places, none of them on the feature list

The first is a single point of failure. One front door means one outage, one quota change, one pricing decision lands on every workflow in the company rather than on one department’s tool. Futurum records that most of these announcements carried no pricing and no general availability date (same analysis). A commitment without a price and a date is a budget that cannot be closed and a risk that cannot be priced.

The second is bundling. An entry point pulls budget away from point solutions and compresses their value into a skill entry in a registry. That is not automatically bad for a buyer, but it changes the renewal conversation: the tools you bought per seat start to look like repeat payments for something the layer above already includes.

The third is moving. When memory lives in a private format, exit cost is not switching vendors, it is teaching the organisation’s context to the next system. This is the hardest to quantify and the easiest to miss, because it costs nothing on signing day and gets charged in full at a renewal three years later.

One more thing gets missed. Once a coworker agent has an email address, a calendar and storage, it is a record in the corporate directory. The problem shifts from deploying a tool to onboarding and offboarding a role: who owns its files and conversations when it leaves, how long its historical actions stay in the audit trail. Technical proposals rarely mention this. Compliance reviews always ask.

The skills registry opens one more permission surface. By design, teams can publish their own workflows into a company-wide registry that other departments’ agents can call (announcement). The blast radius of one publication therefore exceeds the publishing team: once another department calls that skill, a reworded step changes their behaviour too. Someone has to decide who may publish to the company registry, how a publication is rolled back, and who owns the incident when it goes wrong. That rule is rarely in the platform’s defaults.

Four questions for procurement, each with a verifiable answer

First, can memory be exported in a readable form. Not as logs, and not as a summary, but as a structure the next person or system can use directly. The test is cheap: ask for one export, hand it to someone who was not on the project, and see whether they can name the constraints that were in the context.

Second, are skills stored as plain text. If the definition of how a job is done exists as prompts and specifications, it can run somewhere else. If it is compiled into a black box, what you bought is an experience that only executes here.

Third, can a single agent’s identity be revoked on its own, and after revocation can its past actions still be traced to it. This decides whether an incident costs you one surface or the whole front door.

Fourth, what does the spend cap mean. Pause, degrade or keep running, who gets notified, who can change the number. A cap without defined semantics is not a cap in a real budget.

Our judgment

What enterprises are buying this year is not a smarter assistant. It is ownership of organisational memory. Price it by the cost of moving, not by whether the model can be swapped, because the vendor has already decoupled the model for you; whatever cannot be decoupled is where negotiating power actually sits. The judgment carries a testable consequence. If memory and skills export cleanly, moving the same workflow to another platform should mean rewriting an adapter, not re-deriving the acceptance criteria. If every migration forces you to relabel a batch of acceptance criteria, you did not buy a tool. You bought a dependency.

The judgment has boundaries. For a team running single-step tasks at low volume with no cross-system writes, the governance gain from consolidation outweighs the risk of moving, and building an orchestration layer yourself to preserve portability is wasted work. For a company whose workflows cross regulatory boundaries and whose audit records go to an external body, portability outranks integration. Deloitte’s survey of nearly 550 US cross-industry leaders found 80 percent consider themselves mature at basic automation, 28 percent say the same about automation plus agents, and only 12 percent expect the agent work to return its investment within three years (Deloitte TMT Predictions 2026). Companies below that maturity line will save effort by consolidating first. Companies already running critical processes should put portability into the contract before they do.

One thing you can do this week

Pick three real tasks, ideally cross-system ones such as extracting payment terms from a contract and writing them into the ERP. Ask the vendor to export the execution trace for each. Then read the exports together and check what is inside: tool calls, decision points, timestamps for human approval. If the export is a well-written summary, you bought conclusions rather than records. Conclusions cannot be replayed and cannot be moved.

The exercise costs one meeting. What it buys is knowing the exit price before you sign.