Three announcements landed on October 27, 2026, and they describe the same problem from three different angles. Enterprises can build AI agents. They cannot yet say who owns one, what it is allowed to touch, or what it did at 3 a.m. on a Sunday.
IBM used TechXchange in Atlanta to ship agent identity and control in watsonx Orchestrate. Google Cloud's Moe Abdula talked to BetaNews about why pilots stall. Dell expanded its AI Factory with NVIDIA and now claims 6,500 customers, up from 5,000 last quarter. Read together, they sketch the fight over where the agent control plane lives.
IBM puts identity and a kill switch in Orchestrate
IBM's premise is that you cannot govern an agent you do not know exists. Orchestrate's Discover capability (tech preview) connects to Amazon Bedrock, Microsoft Foundry and Google Gemini to enumerate agents built elsewhere, plus models and MCP servers. That matters because most large companies already have agents scattered across three or four clouds and a few laptops.
Contain (also tech preview) runs agents in isolated sandboxes so one misbehaving agent cannot corrupt the rest, and adds a kill switch. Evaluate lets teams test quality, accuracy and compliance before deployment, then monitor in production with per-agent tracking of model and token use.
The identity piece is the part I find most functional. Orchestrate hooks into Vault's credential controls and Agent Identity so each agent gets a defined job, bounded authority and least-privilege credentials that can be revoked when risk changes. That is the correct shape for this problem. An agent without a revocable credential is just an unmanaged API key with opinions.
Whether the cross-vendor discovery actually holds up in a customer environment with 400 agents is an open question. Tech preview means IBM knows it too.
Google argues for one roof, not four
Abdula's Q&A is the most interesting of the three because it disagrees with IBM on architecture. His framing: in Southeast Asia the bottleneck moving from pilot to production is not model capability. It is context and integration. Fragmented markets, strict data sovereignty rules, multilingual nuance and low-latency requirements that a generic sandboxed model does not handle.
The Singapore Engineering Center exists so Google can fold those constraints into the core platform instead of leaving customers to patch around them.
Gemini Enterprise is his answer, and the pitch is a single governed foundation where data permissions, agent lifecycle and cost observability live under one architectural roof. The analogy he uses is a new hire. You do not hand them keys to the whole building on day one. They get the data they need, their actions are logged, and high-stakes decisions need a manager's sign-off. Gemini Enterprise makes agents inherit the strict permissions of the requesting user, leaves an audit trail, and requires human approval before critical actions.
He also names the failure mode most pilots hit: token shock. Unit economics that do not scale predictably against business return. His three success indicators are sound unit economics, sustained daily frontline adoption, and resilience inside existing security boundaries. Without positive economics and clear business ownership, he says, a deployment is an experiment, not a production asset.
I would push back on one thing. The single-roof model is easier to govern, but it assumes the customer consolidates onto one vendor. IBM's bet is that they will not.
Dell sells the metal underneath
Dell's update is the least glamorous and probably the most consequential for anyone running inference at scale. The AI Factory with NVIDIA now validates Reflection's Beam, an Apache 2.0 open-weight model aimed at software engineering and long-horizon agentic workloads. New partners include Dynatrace for AI observability, Mistral for knowledge assistants, and Fogsphere for computer vision, with SUSE Linux support added to the OS list.
The security story is hardware-rooted: NVIDIA OpenShell Reference Architecture plus Red Hat AI Factory software for deploying autonomous agents at scale.
The hardware numbers are specific enough to be useful. The PowerEdge M9882 puts 256 NVIDIA Vera CPUs in a single rack for agent orchestration and code execution. The M8822 is an air-cooled server tuned for agentic workloads, analytics and reinforcement learning. The XE9810 uses Groq 3 LPUs for rack-scale inference. There is a containerized Modular Data Center with NVIDIA DSX POD that ships with compute, storage, networking, power and liquid cooling already integrated, which is the kind of thing you buy when the datacenter team has stopped answering email.
On the deskside side, Dell is pushing agentic AI to developer workstations with templates covering agentic coding, NVIDIA NemoClaw deployments and Hermes Agent implementations, scaling from one box to more than 1,000 systems. The Pro Precision with GB10 gets a 64GB unified-memory option. The GB300 workstation will support Windows with up to 748GB of coherent memory and 20 petaflops of FP4 compute.
Dell's argument is that enterprises want to own the infrastructure and are looking for alternatives to cloud-only deployment. Given what cloud token bills look like at agent scale, that argument has teeth.
What to actually take away
If you are deploying agents this quarter, three things from these announcements are worth stealing regardless of vendor.
Give every agent a revocable credential, not a shared service account. IBM is right that identity is the prerequisite. Google is right that the agent should inherit the requesting user's permissions rather than a superset.
Track cost per agent, not cost per model. Token spend inside an agent loop is invisible until finance asks. Both IBM and Google now expose this. Build it into your own stack if you are rolling your own.
Test the kill path before you need it. IBM's sandboxing and kill switch are the right defaults. A rogue agent is not a hypothetical, and the team that discovers its containment procedure during an incident has already lost the afternoon.
The three vendors disagree about where control should live. IBM wants a layer above everyone's clouds. Google wants one platform. Dell wants you to own the rack. All three can be true in the same company, which is exactly why nobody has a clean answer yet.