Everyone sells the same enterprise agent now. Buy the boring part.

Between September 28 and October 8, four of the biggest names in AI each announced an enterprise agent. Meta launched its Enterprise Platform. xAI opened Team Bots to business plans. OpenAI shipped Dots at DevDay. Google announced the Gemini agent at Gemini at Work.

I read all four announcements back to back, which I don’t recommend as an evening. Somewhere around the third one I started hiding the logo and trying to guess the vendor, and I kept getting it wrong.

They all describe an agent that runs on its own cloud computer, keeps working after you close the tab, signs into your apps through a catalog of connectors, and stops to ask before it sends, buys or deletes anything. Most of them now give the agent its own identity, too. The category converged, and faster than I expected.

If you’re a CIO trying to pick one, comparing the agents won’t get you far. What follows is what I’d compare instead, and what the companies already running these things can actually show for it.

Here’s the Thing

Meta, xAI, OpenAI and Google each announced an enterprise agent within 11 days (Sept 28 to Oct 8, 2026), and they’re architecturally almost identical: a persistent agent with its own VM, app connectors, and approval gates on risky actions.

The real differences are governance: per-agent identity, audit, whether the sandbox is a real security boundary, spend caps, and who owns the registry your agents get listed in.

The adoption stories that hold up are narrow, human-gated, and mostly built in-house. Almost none are both fully autonomous and independently measured.

The agent became a commodity, again

Back in June I argued that in a self-service agent platform, the agents are the easy part. Generating one was already a commodity, and everything wrapped around it was the real work. I meant that about internal platforms. I didn’t expect the frontier labs to prove it about their own flagship products four months later.

Google’s Gemini agent takes objectives rather than instructions, spawns sub-agents for jobs that run “hours or days,” and gives coworker agents their own email address, Drive and directory entry (Google Cloud, 2026). OpenAI’s Dots each get their own cloud computer and browser, with rules that allow, require approval for, or block each action (VentureBeat, 2026). xAI’s Grok Bot runs “on a persistent cloud computer” and pauses before sending, publishing, buying and deleting (xAI, 2026). Anthropic’s Managed Agents run hours-long sandboxed sessions that survive disconnects (Anthropic, 2026).

Each of those is a persistent runtime with connectors, an approval step and an identity. The plumbing between them settled too. MCP is close to universal, and A2A joined the Linux Foundation’s Agentic AI Foundation in August with more than 150 backers (A2A Project, 2026).

A converged category is easier to buy. It also means the keynote demo tells you almost nothing, because every vendor is now giving the same demo.

Meta and xAI, honestly

These were the two I was most curious about, because they’re the newest to enterprise and the loudest about it. Both are shipping real software, and neither is ready for the job its announcement implies.

xAI

xAI ships fast, and I’ll give it credit for unusually honest documentation. Grok Bot launched in beta in August, and Team Bots followed at the end of September with shared memory, admin assignment, and connector policies (xAI, 2026).

Read those docs, though. They warn, twice, that separate bots are not separate security sandboxes. Any secret you give a team bot is usable by every teammate, and Slack channels share one computer. Audit logs are Enterprise-only. The plugin catalog has around 220 entries and no HR or IT service management tools (Reworked, 2026). I couldn’t find one named commercial customer for either product. The named customers are US government agencies.

Grok Bot is a beta, and xAI says so. It only becomes a problem if you buy it as something more finished.

I use Grok personally, not at work, and I like it. It’s very proactive. Where it breaks for me is scheduled routines, which can fail silently. A schedule that might not run, and won’t tell you when it didn’t, is not something I’d put a business process on.

Where they actually differ: the boring part

Once the agent is a commodity, the differentiation moves to everything around it. Much of that is territory I spent most of my time in building AgentHive: identity, audit, permissions, the catalog. The vendors have added sandboxing and cost control to the list.

On paper, Google is furthest along. Each agent gets an identity Google calls “cryptographically attested,” the audit trail attributes actions to the agent rather than a person, every agent runs in its own sandbox with its own network boundary, and hard spend caps pause the agent when it hits its budget (Google Cloud, 2026). It’s also in private preview with no pricing, so “on paper” is doing real work in that sentence.

OpenAI has the right architecture and an availability problem. Frontier, announced in February with HP, Intuit, State Farm and Uber as first customers, gives every agent its own identity and scoped permissions (TechCrunch, 2026). Eight months later I can’t find a general availability announcement. OpenAI also published something I wish every vendor did: Dots passed 45 of 49 tests on halting when permissions change mid-task (BankInfoSecurity, 2026). Four failures out of 49 is a number you can plan around, and it’s the only published failure rate I found from any of these vendors.

Anthropic has shipped the most. Cowork is generally available with role-based access and spend limits, and Managed Agents now support sandboxes that execute inside your own perimeter (Anthropic, 2026). The caveat is procurement: the Pentagon’s supply chain risk designation was upheld on appeal in September. That matters if you’re a defense contractor and doesn’t if you aren’t.

The registry fight is the part I’d watch most closely. Microsoft, Salesforce, ServiceNow, AWS, SAP and Workday each now ship an agent registry, and each one is built to discover the other vendors’ agents too. Microsoft says Agent 365 has “nearly 40 million agents registered across tens of thousands of companies,” two months after launch (Microsoft, 2026). Salesforce’s Agent Fabric scans Bedrock, Foundry, Vertex and Copilot Studio. ServiceNow’s AI Control Tower added 30 integrations, including SAP, Oracle and Workday.

Everyone wants to be the place your agents are listed, because whoever holds the system of record holds the renewal. I said in June that the control plane is “a decision you make once and regret slowly,” and every vendor in this list has clearly read the same memo.

Pricing is shifting underneath all of it. Salesforce charges $0.10 per agent action or $2 per resolution (Salesforce, 2026). Atlassian starts billing Rovo overage in December, on by default. ServiceNow, SAP and Workday meter in units whose price they don’t publish. Seats were predictable. Actions aren’t, and an agent stuck in a loop now shows up on an invoice.

What the adoption evidence actually says

The vendor numbers are big. The case studies underneath them are more interesting, and more mixed.

In engineering, the strongest evidence comes from companies that built their own background agents. Stripe’s “Minions” turn a Slack message into a pull request that passes CI, and merge more than 1,300 PRs a week, with a human reviewing every one (Stripe, 2026). Shopify’s River co-authors one in eight merged PRs (Shopify, 2026). Spotify took a 1,000-PR migration from three months to about ten days, and also reported that its agents occasionally “cheated” by commenting out tests (InfoQ, 2026). I trust that last detail more than any headline number, because nobody puts it in a sales deck.

The famous examples are thinner than their reputations. Goldman Sachs’s Devin pilot made headlines in 2025, but I couldn’t find a published outcome for it. Nubank’s “12x efficiency, 20x cost savings” figures come from Cognition’s customer page, which sells Devin, while Nubank’s own engineering blog gives no numbers (Devin, 2024). Google says 75% of its new code is AI-generated and engineer-approved, without saying what it’s counting (Google, 2026).

Outside engineering, almost every success story has a human approving the consequential step. Allianz Australia’s seven-agent claims system cut processing time for small claims by 80%, and payouts are “never automated” (Allianz, 2025). Salesforce cut its own support headcount from about 9,000 to 5,000 with agents handling roughly half of conversations (Fortune, 2025). Klarna’s assistant did the work of 853 people, and Klarna’s CEO still called the result “lower quality” and started putting humans back (CX Dive, 2025). Commonwealth Bank cut 45 call centre roles for a voice bot, watched call volumes go up, and reversed the decision within weeks (ABC, 2025).

Then there’s the quality side of the ledger, which nobody publishes. Faros AI’s telemetry across 22,000 developers found merged PRs up, along with incidents per PR up 242% and review time up 441% (Faros AI, 2026). It’s correlation from a vendor, so hold it loosely. I went looking for a single company that publishes the revert rate on its agent-authored PRs, and the closest I found was Shopify saying it’s “roughly the same.”

The quadrant nobody is in

If I took one thing from all this into a vendor meeting, it would be this sort.

Put every case study on two axes. One is autonomy: is the agent assisting a person, acting with a human approving, or acting fully on its own? The other is evidence: did the company report the result, did the vendor, or did someone independent measure it?

Almost everything lands in the bottom two rows, reported by the company running the agent or the vendor selling it. The top row, where someone independent measured the result, has two entries. One is METR’s trial of experienced developers working with AI assistants, which found them 19% slower (METR, 2025). The other is Commonwealth Bank’s voice bot, the only fully autonomous case I found with independent reporting, and it was reversed within weeks. The analyst numbers point the same way: McKinsey found 23% of companies scaling an agentic system anywhere, and no more than 10% doing it in any single function (McKinsey, 2025).

The agents that work share a shape: narrow scope, a human on anything irreversible, and usually built by the company running them. The products the labs just announced are designed for far more autonomy than that, and the evidence that it works at that level doesn’t exist yet, though it might by next year.

What I’d buy on

If I were evaluating one of these tomorrow, the agent demo would get five minutes. These questions would get the rest of the hour.

Does every agent have its own identity, and does it die when its owner leaves?

Per-agent identity is table stakes now. Wiring it to offboarding isn’t. I wrote about what happens when nobody is on call for what employees built, and agents with their own email accounts make that problem bigger.

Is the audit log in the plan I’m buying, or an upsell?

If you need Enterprise to see what an agent did, price it as Enterprise from day one.

Is the sandbox a security boundary?

Ask directly. xAI’s docs say no, which at least is honest. If a vendor won’t answer, treat that as a no.

What stops a looping agent from spending money?

Look for hard caps that pause the agent at its budget. An alert email arrives after the money is gone.

Whose registry is this, and can I leave?

If the registry only lives in one vendor’s cloud, what you’re buying is that vendor’s gravity.

What’s your number for when it fails?

Revert rate, escalation rate, the halting tests, and whether a scheduled run that fails tells anyone. OpenAI published 45 of 49. Ask everyone else for theirs.

Frequently asked questions

What are Meta and xAI offering enterprises in 2026?

Meta’s live product is Business Agent for customer messaging on WhatsApp, Messenger and Instagram, used by more than a million businesses. Its Enterprise Platform, announced September 28, has no published pricing, controls or customers yet. xAI offers Grok Bot and Team Bots in beta, with audit logs on Enterprise only and government agencies as its only named customers.

What do enterprise agent platforms have in common now?

Nearly all run each agent on a persistent cloud computer, connect to business apps through connectors and MCP, pause for approval before risky actions, and assign the agent its own identity. Google, OpenAI, Anthropic and xAI all describe this architecture. They differ in governance maturity and general availability.

Are companies actually running autonomous agents in production?

Yes, but mostly with a human approving the consequential step. Stripe merges more than 1,300 agent-written PRs a week with human review on each. Allianz automates claims processing but never payouts. McKinsey found only 23% of companies scaling an agentic system anywhere, and no more than 10% in any single function.

What should a CIO evaluate when buying an agent platform?

Per-agent identity tied to offboarding, audit logs included in the plan, a sandbox that’s a real security boundary, hard spend caps, a registry you can leave, and published failure rates. The agent itself is close to a commodity across vendors, so these are where the products differ.

Where I landed

I started this expecting to compare agents and ended up writing about everything around them, which is a little embarrassing given that I made the same argument in June about our own platform and still had to rediscover it.

The labs have done something useful by agreeing on what an enterprise agent is. The next year of this market will be decided by whose identity model, audit trail, sandbox and registry you can live with for five years.

So read the next announcement the way I wish I’d read the last four. Skip the part where the agent books a meeting, scroll to the paragraph about governance, and see whether there’s anything in it.