Orbital / IBM Agent Identity: a business control checklist for AI agents

IBM Agent Identity: a business control checklist for AI agents.

IBM made Agent Identity generally available and announced new cross-platform agent controls. Use this checklist to assess identity, access, evidence and cost.

Agent identity · Business controls

By Orbital Content Team · Published October 27, 2026

AI-assisted research and drafting with editorial review. Release, availability, feature and pricing statements were checked against IBM's official announcements and product pages on October 27, 2026. IBM product claims are identified as vendor-reported. The evaluation framework and business implications are Orbital analysis. Orbital did not use the products, inspect an IBM customer environment, verify cross-platform discovery, test identity delegation or revocation, audit runtime isolation, reproduce performance claims, confirm an Agent Identity price or review customer-specific terms.

IBM announced a broader control layer for business AI agents on October 27, one day after making IBM Agent Identity generally available. The release combines an agent inventory and identity record with delegated authorization, credential controls, cross-platform discovery, evaluation, cost visibility, isolation and shutdown controls across IBM and selected third-party environments.

The important business lesson is larger than one vendor: an agent that can act in a CRM, project system, inbox, database or payment workflow needs a named owner, a bounded identity, limited credentials, verifiable results and a way to stop. A helpful demo does not establish those controls.

What IBM announced on October 26 and 27

IBM's October 26 announcement says IBM Agent Identity is generally available. IBM describes it as an identity control plane that can register agents in a central inventory, record ownership and lifecycle state, preserve on-behalf-of delegation, issue or coordinate short-lived task-scoped access, apply contextual authorization and connect audit evidence across people, agents, credentials, policy decisions and downstream actions.

IBM says the product works with IBM Verify, HashiCorp Vault and watsonx Orchestrate and is designed to coexist with existing identity providers. The related product page names Microsoft Entra ID, Okta and Ping Identity among the surrounding identity ecosystem.

On October 27, IBM announced additional agent-management capabilities:

  • Discover, in tech preview: watsonx Orchestrate is intended to find agents, models and Model Context Protocol servers in selected third-party environments such as Amazon Bedrock, Microsoft Foundry and Google Gemini.
  • Contain, in tech preview: IBM says agents can run in isolated sandboxes and administrators can use a kill switch.
  • Evaluate: the announcement describes predeployment and production evaluation, trace inspection, deterministic journey tests, LLM-as-a-judge criteria and model/token cost tracking.
  • Credential controls: IBM says Vault can provide short-lived, least-privilege credentials and revoke access when an agent attempts an unauthorized action.
  • Catalog and gateway: IBM describes a governed catalog and an AI Gateway for applying observability, evaluation and policy across agents, models and MCP servers.

IBM also labels Secure Hub as a public preview. The announcements mix generally available products, current product capabilities, public previews and tech previews. Buyers should verify the status of every required feature in their account and region rather than treating the release as one uniformly available package.

Separate documented facts from business implications

What the announcement supports—and what it does not establish
AreaDocumented statementPractical implicationNot established
Agent identityIBM Agent Identity is generally available and is designed to register, authorize and govern agents.A buyer can evaluate a released identity product rather than only a roadmap.That it supports every framework, protocol, region, identity provider or workflow the buyer uses.
CredentialsIBM says Agent Identity and Vault can coordinate task-scoped authorization and short-lived credentials.Persistent secrets do not have to be the default for every agent connection.That the target application accepts short-lived credentials or that revocation prevents every unwanted action.
Cross-platform inventoryDiscover is identified as a tech preview for selected external environments.Multi-vendor inventory is a direction worth testing.Production coverage, completeness, support terms or discovery of every shadow agent.
Isolation and shutdownContain is identified as a tech preview and includes sandboxes and a kill switch.Runtime containment can be evaluated as a control outside the prompt.Isolation strength, recovery time, data residue or behavior under every failure mode.
EvaluationIBM describes deterministic tests, user simulation, traces, custom judge criteria and production monitoring.Agent quality can be treated as an operating process rather than a launch checklist.That a judge model is correct, that test coverage is complete or that monitoring prevents customer harm.
InteroperabilityIBM positions Orchestrate as open and multi-platform.A buyer may avoid rebuilding every external agent in one framework.Low switching cost, full feature parity or freedom from commercial and operational lock-in.
PricingIBM publishes watsonx Orchestrate plan prices and asks prospects to contact IBM for Agent Identity.A buyer has a starting point for one layer of the stack.The combined price, implementation cost, required licenses, usage charges or total operating cost.

The minimum viable control model

A small or midsize business may not need a large agent-management platform on day one. It still needs the control outcomes the announcement highlights. Before an agent touches a live system, require:

  1. Identity: a stable agent ID that is distinct from a developer, shared admin or generic integration account.
  2. Ownership: one business owner and one technical owner who can approve changes and disable the workflow.
  3. Purpose: a narrow task, approved systems, expected result and explicit exclusions.
  4. Authority: the person, policy or system that delegated the task and the duration of that delegation.
  5. Least privilege: only the records, tools and actions required for the approved task.
  6. Evidence: traceable inputs, tool calls, approvals, outputs and downstream results without unnecessary sensitive content.
  7. Evaluation: fixed test cases, production signals, thresholds and a review owner.
  8. Containment: limits on destinations, records, actions, time, steps, spend and concurrency.
  9. Stop and recovery: a tested kill switch, credential revocation, rollback where possible and a manual fallback.
  10. Lifecycle: review dates, change control and removal when the agent or integration is retired.

This can begin with existing identity, secrets, logging and workflow tools if they can produce the required control and evidence. A new platform is justified by the gap it closes, not by the number of features in a product announcement.

Use Orbital's AI automation readiness checklist to define the trigger, owner, data boundary, approval and fallback before selecting an agent platform.

Build an agent register before buying a control plane

Create one row for every agent, scheduled AI workflow and tool-using assistant that can reach business data or perform an action. Record:

  • agent name, stable ID, version and lifecycle state;
  • business owner, technical owner and support contact;
  • approved purpose, users and environments;
  • model, framework, hosting location and vendor;
  • connected applications, APIs, MCP servers and data stores;
  • identity provider, service account, credential broker and secret location;
  • permissions, action classes and approval gates;
  • data categories, retention, processing locations and subprocessors;
  • evaluation set, pass thresholds and last test date;
  • logs, traces, alerts and incident runbook;
  • token, tool, platform and support cost owner; and
  • expiration, next review and retirement procedure.

Reconcile the register against identity-provider service principals, secrets stores, cloud accounts, automation platforms, API keys, scheduled jobs and vendor consoles. Discovery software may help, but a tech preview or scan result should not become the only inventory evidence.

Coordinate the register with the business software integration checklist so agent ownership, systems of record, retries and reconciliation are documented together.

Define authority, credentials and action classes

An agent identity is useful only when it changes what the agent can do. For each workflow, write an authority contract:

  • who or what initiated the run;
  • which agent received the task;
  • the approved purpose and time window;
  • the records, fields and destinations it may access;
  • read, draft, reversible write and irreversible action permissions;
  • actions requiring human approval;
  • credential issue, renewal and revocation rules;
  • maximum records, tool calls, duration and spend;
  • how the saved result is verified in the system of record; and
  • what causes an immediate stop.

Do not give an agent a founder's, administrator's or employee's standing credential merely because it is convenient. Prefer a dedicated identity and short-lived, purpose-specific access where the target systems support it. If a legacy application accepts only a persistent secret, isolate it, narrow the account permissions, store it in an approved secrets system, rotate it and record every use.

Keep money, publishing, deletion, access changes, legal commitments and external customer messages behind explicit approval until the business has evidence for a narrower policy. A model instruction such as “ask before sending” is not an enforceable access control.

Require evidence before and after an action

IBM's announcement usefully separates inventory from evaluation. A registered agent can still perform the wrong task correctly, use an incorrect tool parameter or report success when the destination did not save the change.

Create an evaluation set from real workflow variation using sanitized or synthetic records:

  • normal, incomplete, conflicting and out-of-scope requests;
  • wrong customer, duplicate event and stale data cases;
  • permission denial, expired credential and revoked access;
  • rate limit, timeout, unavailable destination and partial failure;
  • malicious or irrelevant instructions in documents, web pages and tool responses;
  • an action that requires approval and one that is prohibited;
  • a changed interface, field or tool schema; and
  • a recovery case after the workflow stops mid-task.

Score the business result, not only the final response. Verify the record that was read or written, the tool and parameters used, the approval captured, the downstream state, duplicate behavior, correction time and total cost. Treat LLM-as-a-judge scores as one signal; maintain deterministic checks and knowledgeable human review for high-impact cases.

In production, monitor accepted-result rate, correction rate, unauthorized attempts, approval overrides, tool failures, duplicate or missing records, queue age, latency, token and tool use, incidents and manual fallback. Set a stop threshold before launch.

Separate generally available controls from previews

The October 27 release uses several availability labels. Preserve them in procurement and pilot documents:

Availability questions for the IBM announcement
CapabilityIBM's stated statusBuyer verification
IBM Agent IdentityGenerally available October 26.Regions, required Verify or Vault components, supported identity providers, APIs, service levels, pricing and contract terms.
Orchestrate DiscoverTech preview.Supported platforms and asset types, inventory completeness, refresh timing, permissions, export and production-support limits.
Orchestrate ContainTech preview.Isolation boundary, supported runtimes, shutdown behavior, evidence, recovery and preview data restrictions.
Secure HubPublic preview.Environment coverage, production use terms, support, data handling and transition plan.
Evaluation and cost trackingDescribed as current Orchestrate capabilities.Plan entitlement, supported external agents and models, metric definitions, export, retention and added usage cost.

A preview can be appropriate for a non-production evaluation. It should not silently become a dependency for a customer-impacting production control. Require a fallback if the preview changes, ends or does not cover the intended environment.

Decide whether an agent-control platform fits

Signals for a lighter control stack or a dedicated platform
SituationLikely starting pointDecision evidence
One low-risk internal agent with read-only accessExisting identity, secrets, logging and approval tools may be sufficient.Can the business identify, limit, trace, stop and review it without gaps?
Several agents in one vendor ecosystemEvaluate the native control plane first.Coverage, export, policy consistency, cost and portability.
Agents across multiple clouds and frameworksA cross-platform inventory and policy layer may reduce fragmented controls.Demonstrated discovery completeness and enforcement for the actual environments.
High-impact writes or regulated dataFormal identity, secrets, evaluation, isolation and audit architecture.Independent security, privacy, legal and compliance review plus acceptance tests.
No stable process or system of recordFix the workflow before adding an agent platform.Named owner, approved process, clean data and measurable outcome.

Do not buy governance software to avoid making governance decisions. The platform can enforce and evidence a policy, but the business must still decide who owns an agent, what authority is appropriate, which outcomes are acceptable and who responds when it fails.

Model complete cost—not one plan price

IBM's public watsonx Orchestrate pricing page listed a 30-day trial, Essentials starting at $530 per month, Standard starting at $6,360 per month and Premium as contact-based pricing when checked October 27. The same page lists different capacity and isolation features by plan. Those prices concern Orchestrate; the Agent Identity product page directs buyers to book a demo and did not provide a numeric price in the evidence reviewed.

Build a cost ledger that separates:

  • Agent Identity, Verify, Vault, Orchestrate and required platform licenses;
  • models, tokens, tools, MCP services, connectors and third-party APIs;
  • cloud runtime, storage, network, logs and trace retention;
  • implementation, identity integration, policy design and data mapping;
  • evaluation labels, test environments and release review;
  • security, privacy, legal, procurement and audit work;
  • monitoring, incident response, credential rotation and upgrades;
  • human approvals, corrections and manual fallback;
  • preview migration, vendor change and exit costs; and
  • the expected cost of an incorrect accepted action.

Ask for a line-item quote using the expected agent count, active users, messages, tool calls, connected environments, regions, log retention and support level. A starting subscription price does not establish total cost or business value.

Ask vendors to prove the control path

Give every provider the same bounded workflow and request evidence for:

  • supported agent frameworks, identity providers, clouds, MCP servers and applications;
  • inventory discovery method, frequency, blind spots and export;
  • human-to-agent and agent-to-agent delegation across token exchanges;
  • credential issue, scope, lifetime, renewal, revocation and audit;
  • policy enforcement before and during tool use;
  • isolation boundary, kill-switch behavior and recovery;
  • deterministic tests, traces, judge models, feedback and production monitoring;
  • data categories, retention, processing regions, subprocessors and customer controls;
  • preview versus generally available features and production support;
  • availability, incident response, service limits and change notice;
  • complete pricing, required products, implementation and support; and
  • data, policy and audit-log export plus an exit plan.

Request a live demonstration using a test account: issue a narrow task, approve it, show the delegated identity, obtain a short-lived credential, attempt one prohibited action, revoke the session, inspect the trace, verify the destination record and export the evidence. A slide describing least privilege is not proof of enforcement.

A bounded four-week pilot

  1. Week 1 — inventory and scope. Select one reversible internal workflow. Record the agent, owners, systems, data, authority, actions, current baseline, fallback and stop conditions.
  2. Week 2 — control design. Create the dedicated identity, least-privilege permissions, credential path, approval gate, logging, limits and evaluation set. Confirm feature status and terms.
  3. Week 3 — non-production tests. Run normal, prohibited, duplicate, stale-data, outage, revoked-credential and recovery cases. Verify downstream state and evidence, not just agent text.
  4. Week 4 — shadow and decide. Compare the agent's proposed work with the existing process without allowing independent customer-impacting action. Measure accepted results, corrections, failures, review time, complete cost and control gaps.

End with a recorded stop, revise, continue or expand decision. Expand one dimension at a time: action, system, data class, user group or volume. Do not add autonomous customer messages, payments and access administration in the same release.

Orbital's business take

IBM's announcement is useful because it moves the conversation from “what can the agent do?” to “whose authority is it using, what credential does it hold, what can stop it and what evidence remains?” Those questions apply even when a business never buys IBM software.

For most New York service businesses, the first step is not an enterprise control plane. It is one trustworthy register, one bounded workflow, dedicated credentials, explicit approval for customer-impacting actions, outcome verification and a manual fallback. If multiple agents and vendors make those controls fragmented, evaluate a platform against a written control gap and a realistic test—not a generic agent demo.

Use Orbital's AI automation services, tech stack consulting, connected-business planning, business consulting and service-business planning to define the operating model before implementation. The AI automation ROI checklist can connect control costs and failure risk to a stop-or-scale decision. A 60-minute consultation can turn one workflow into an evaluation brief; software, implementation, usage, security, privacy, legal review and ongoing support are separately scoped.

Sources and editorial limits

Primary IBM sources were checked October 27, 2026. IBM's October 27 newsroom release and watsonx Orchestrate announcement describe vendor capabilities and roadmap status; they are not independent evaluations. The October 26 Agent Identity announcement identifies the product as generally available. IBM's product and pricing pages were used for current positioning and published Orchestrate plan prices. The reviewed Agent Identity page did not provide a numeric price. Orbital did not use the products, inspect a customer environment, verify discovery completeness, test delegation or revocation, audit runtime isolation, reproduce performance claims, review customer-specific licensing or determine legal, privacy, security or regulatory compliance. No feature or control guarantees security, accuracy, cost savings or business results.

Primary source links

Your next chapter.
Let’s give it lift.

Start a conversation