Content
The IBM TechXchange 2026 announcement enterprise leaders should not miss gives every AI agent its own identity, with a named owner, limited access and a record of what it did.
IBM announced the package on October 27, the first full day of the conference in Atlanta. You can buy three parts today: IBM Agent Identity, HashiCorp Vault 2.2 and new testing tools in watsonx Orchestrate. The tools that find agents on other platforms, isolate them and switch them off are still in tech preview.
IBM’s own summary of the problem is blunt: “Enterprises are deploying agents faster than they can verify them.”
Before the event, our TechXchange preview asked who owns each agent, what it can touch and how fast you can stop it. IBM now ships an answer to the first two questions. The third is still in preview.
What IBM announced at TechXchange 2026
Ask most IT leaders how many AI agents run in their company and who approved each one, and you will get an estimate. Gartner found that only 13% of organizations think their agent governance is adequate. IBM’s package targets that gap, with pieces spread across IBM and HashiCorp products and shipping on different timelines:
| Component | What it does |
|---|---|
| IBM Agent Identity | Keeps one record per agent: owner, lifecycle status, approved tools and granted authority. Available now. |
| HashiCorp Vault 2.2 | Gives agents their own secured accounts and renews their credentials automatically. Available now. |
| watsonx Orchestrate – evaluate | Tests agents, models and MCP servers before go-live, then monitors them in production, with cost tracking. Available now. |
| watsonx Orchestrate – discover and contain | Finds agents built on other platforms, runs them in isolated sandboxes and adds a kill switch. Still in preview. |
| IBM Integration Intelligence Hub | Gives AI assistants context, approved runbooks and policy controls for integration work. Still in preview, launching first as SaaS. |
How IBM Agent Identity works
The rest of the package leans on IBM Agent Identity, so it deserves the closest read. IBM breaks it into four steps:
- Register. Every agent gets an entry in a central inventory, with an owner, a risk class, approved tool connections and a lifecycle state.
- Delegate. An agent acting for a person receives short-lived OAuth tokens limited to that one task. The person’s identity travels with the request.
- Enforce. Access is checked while the agent runs, based on context, policy and risk.
- Audit. One trail connects the person, the agent, the credentials it received, the policy decision and the resulting action.
For systems that can’t accept delegated tokens, Vault issues just-in-time credentials, which keeps long-lived passwords out of agent code. IBM says Orchestrate now connects directly to IBM Verify and Microsoft Entra.
The release lists well-known Vault users such as BlackRock, Target and London Stock Exchange. Those are Vault customers in general. IBM has not named anyone running the full agent identity setup in production yet.
Giving AI agents access to your core systems?
We modernize MQ, ACE, DataPower and mainframe integration in stages, so agents reach core systems through controlled, auditable paths while daily transactions keep running.
Start with one call, no commitment.
Start with one call, no commitment.
What changed and what did not
A vendor announcement can sound finished when half of it is still a preview, and that gap is where budgets get burned. The firm change here is status: Agent Identity was a private preview in September and is now generally available, with IBM support and a normal procurement path.
We went into the event with five checks for IBM’s demos. This is how the announcement scores against them:
| What we asked for | IBM’s answer |
|---|---|
| An inventory with a named owner for every agent | Agent Identity registry, available now |
| User identity carried through each agent call | Task-scoped delegated tokens, available now |
| Policy checks before an action runs | Runtime authorization, available now; human approval for high-risk integration actions, still in preview |
| Traces an auditor could follow | Audit trail from person to agent to action, available now |
| A way to stop an agent and undo what it did | Kill switch in Orchestrate, still in preview; no rollback feature announced |
Four checks have an answer you can buy today. The one still open, stopping and undoing, matters most to a regulated business.
What stays the same
- Controls are spread across products. Policy now lives in Agent Identity, Verify, Vault, Orchestrate, watsonx.governance and DataPower, and IBM hasn’t said which one wins when two disagree.
- Non-IBM agents are only partly covered. Discovery and containment for agents on other platforms are still in preview.
- Undo is still on you. A kill switch stops the agent, but the orders, records or payments it already changed stay changed. Reversing them takes retry rules, duplicate checks and compensating steps in your integration layer.
Enterprise impact: data, integration, governance and cost
Every agent that touches an ERP, a CRM or a core banking system is a new account that an auditor will ask about. Identity decides whether an agent gets near those systems at all, which is why this announcement lands well beyond the security team.
Data: a verified agent can still read bad data
An identity check tells you who acted. It tells you nothing about whether the customer record behind the action was current or correctly defined. A fully authorized agent working from stale data still makes the wrong call, so data readiness stays a separate piece of work.
IBM had data news too. On October 21 it made agentic data integration generally available in watsonx.data integration: describe the dataset you need in plain language, the agent proposes a pipeline, and a person reviews it before deployment. IBM’s documentation currently limits it to DataStage flows on IBM Cloud in selected regions. An agent that builds pipelines is also one more agent to register.
Integration: identity has to survive the middleware
Delegated tokens work in modern cloud apps. They tend to break in the middle layer, where message queues, ESBs, API gateways and core systems often know only shared service accounts. Vault’s just-in-time credentials cover part of that gap. Tracing every hop from the agent to the core system is still integration work.
Integration Intelligence Hub is IBM’s attempt to close it. Agent actions on integration assets go through DataPower Interact Gateway, which checks identity and access, logs the action and can require a person to approve risky steps. It starts as SaaS, so companies running MQ, App Connect or DataPower on-premises should expect to wait. More on the patterns in Agent Sprawl Is an Integration Problem.
Governance: evidence an auditor can follow
For AI governance, the audit trail is the most useful piece. A record that connects a person, an agent, a credential, a policy decision and an action holds up in an audit far better than a policy PDF.
Somebody still has to own it. Decide which product holds the authoritative policy and who signs off each agent’s permissions. Our guide to AI governance for regulated enterprises walks through that operating model.
Cost: no price tag yet
IBM published no prices for the package. Orchestrate’s cost tracking will at least show what each agent spends on models and tools. The bigger bill is likely to be the work around the licenses: connecting Agent Identity to each target system, retiring shared service accounts, rolling out Vault, and licensing four or five IBM products where you may have budgeted for one.
Who should act now, and what to verify first
Buy too early and you pay for previews. Wait too long and agents keep running on shared passwords nobody owns. Our preview set a strict bar for deployable now: general availability, documentation, support and customers in production. Nothing here clears it yet, because IBM has not named a customer running these pieces together. Most of the package is still a strong pilot candidate.
Pilot with controls: start next quarter
- You already run IBM Verify and HashiCorp Vault. Registering agents in Agent Identity and moving them off shared service accounts is a contained first step in AI agent identity management.
- You have agents in production in a regulated business. Test the registry and audit trail before an auditor asks who approved an agent and what it did.
- You build agents in watsonx Orchestrate. The evaluation tools carry little risk, and testing before go-live should be routine anyway.
Watch for now
- Your agents run on Amazon Bedrock, Microsoft Foundry or Google Gemini. Discovery and containment for those platforms are still in preview.
- Your integration estate is mostly on-premises. Design the target state now, and commit once IBM confirms on-premises support for Integration Intelligence Hub.
- Your plan depends on a kill switch. Until containment is generally available, keep your own emergency stops: revocable credentials, feature flags and rate limits.
Five questions to ask IBM before you buy
Handle AI agent security the way you would any non-human identity program, and confirm the scope before you scale.
- Which deployment models does Agent Identity support today: SaaS only, or also on-premises and air-gapped?
- How does a delegated user identity pass through MQ, API gateways and core systems such as IBM Z?
- When will containment and the kill switch be generally available, and will they cover non-IBM agents?
- Which product holds the final policy when Agent Identity, Orchestrate and watsonx.governance disagree?
- What is the total cost per agent across Verify, Vault, Orchestrate and governance licenses?
Key takeaways
- IBM’s main TechXchange 2026 announcement gives AI agents a registered identity, task-scoped credentials and an audit trail, and those parts are available now.
- The kill switch, discovery of non-IBM agents and Integration Intelligence Hub are still in preview. Nothing reverses what an agent already changed.
- No pricing and no named customer yet, so plan a pilot before a rollout.
- Identity has to hold across gateways, queues and core systems, whichever vendor you choose.
- You can start without buying anything: list your agents, give each an owner and retire shared service accounts.
Our recap of the rest of IBM TechXchange 2026 follows next week.
