IBM Power Autonomous Operations, Explained Before GA: An AI Agent That Runs the Box, Not the Code

IBM Power Autonomous Operations is an AI agent that runs the machine, not the code on it. It watches your Power systems, analyzes what it finds, proposes a fix, and resolves capacity constraints once a person approves, and you drive it with chat-style prompts instead of command syntax. IBM announced it on July 15, 2026 and set general availability for September 23, 2026, which is three weeks from today.

Most of the confusion about this product comes from the calendar. IBM announced it on the same day as the Bob Premium Package for i, and both got filed under the same headline shape: IBM ships AI for IBM i. They are not the same product, they do not sit at the same layer, and only one of them was built for your developers.

What IBM has actually said it does

Three steps, in IBM's own description. The agent continuously monitors Power system behavior and takes in the alerts an administrator would otherwise catch by checking a dashboard. It runs AI-driven diagnostic analysis on what it sees. Then it produces a recommended remedial action, with human-in-the-loop approval before anything is remediated.

That clause is IBM's own, and it matters. This is not a product that quietly rebalances your memory pools at 2 a.m. and tells you in the morning. A person still says yes.

The other half of the pitch is the interface. IBM's framing is that a team should manage, tune and streamline a Power environment through ordinary conversation, without deep domain expertise for every task. Hillery Hunter, General Manager for IBM Power and CTO of IBM Infrastructure, described the goal as "making Power increasingly self-operating."

One thing IBM did not say anywhere in that announcement: IBM i. Power Autonomous Operations is described against the Power environment, not against a partition operating system. IBM i appears in the same release in a different section, attached to Bob. That is not evidence against IBM i support, because capacity pressure at the Power layer is felt by every partition on the box no matter what runs inside it. It is a reason not to assume.

The 15x number, and what actually produced it

The marketing line is that it resolves capacity constraint issues up to 15x faster than manual intervention. The receipt behind it is a single internal test, and IBM published the actual minutes rather than only the multiple, which is worth acknowledging.

Read it the way it was produced. One fault class, capacity-related conditions. One environment, eleven systems, run by IBM. One comparison, against a manual workflow IBM has not fully described. It is a narrow number that is at least honest about its own units.

Also worth noticing: 52.59 minutes is not a bad manual result. That is the number for a team that already knows what it is doing. The value here is not rescuing an outage at 3 a.m., it is handing back the hour burned on a condition nobody should have had to chase. Multiply that by however many capacity events your environment actually produces in a year and you have the real business case, which is a number only you can produce.

Bob writes the code. This one runs the box.

IBM Bob is a development tool. A developer points it at RPG or COBOL and asks it to explain a program nobody has touched since 2004, convert fixed-format to free-format, generate a module or write the unit tests. Since the June 24, 2026 Premium Package for i it does that on source where it lives, in QSYS and the IFS. It has no opinion about your LPAR weights. For what shipped and what it costs, see IBM Bob explained and the IBM Bob directory listing.

Power Autonomous Operations runs the other direction. It has no opinion about your RPG. It watches the machine underneath every partition you run, and its published use case is capacity, not code.

In a large enterprise those are two different teams with two different budgets. In a typical IBM i shop they are the same person: whoever understands the HMC, the shared processor pool and the memory weights is usually carrying the RPG backlog too, and infrastructure work gets whatever hours are left. That overlap is the bottleneck this product aims at, and it is why a conversational interface is more than a demo trick. Approving a capacity fix should not require the one specialist who is on vacation.

It is not watsonx Orchestrate either

watsonx Orchestrate is the business-workflow side of IBM's agent story. It hosts agents that automate multi-step work across business applications, and since the August 2026 release it can launch Bob from the Create an agent dialog to build those agents, an option IBM's release notes list as a preview on IBM Cloud and AWS rather than as generally available. Orchestrate still ships no IBM i connector. See the watsonx Orchestrate listing and watsonx on IBM i for where that stands.

Say the three jobs out loud and they sort themselves. Orchestrate automates a business process. Bob writes software. Power Autonomous Operations keeps the infrastructure healthy. Nothing IBM has published connects this product to the watsonx brand at all, which tells you something: it came out of the Power organization, not the AI platform organization.

What an IBM i shop should ask before this goes in a budget

Here is the honest inventory three weeks out. The first column is the question a buyer has to answer. The second is what IBM has published as of today.

None of that is a knock on the product. Pre-GA announcements are thin by design, and IBM was more specific than usual by publishing a measured figure instead of an unsourced percentage. But "IBM has not published it" is the right answer to give your CFO this month, and that table is a good agenda for the first conversation with IBM or your Business Partner.

The practical position for an IBM i shop today: do not budget for it, do write down which capacity conditions currently cost you a person and an afternoon, and compare that list against what IBM documents at GA. One or two Power servers with a handful of partitions and no chronic capacity pressure makes this interesting rather than urgent. If you are running eleven systems, IBM already built the test case for you.

What to watch on September 23

Four things will tell you how seriously to take this by the end of the month:

• Whether IBM publishes IBM i-specific documentation, or leaves the whole thing Power-generic.
• Whether the announcement letter finally names integration points: HMC, PowerVC, the Cloud Management Console, or something new.
• Whether licensing arrives as its own product, folds into an existing Power software bundle, or lands as a subscription per system.
• Whether capacity constraints are still the only published use case a month after GA. A second use case would say the agent generalizes. Silence would say it does not, yet.

IBM TechXchange runs October 26 to 29 in Atlanta, the first venue where sessions could put real detail behind the launch, so the picture should be clearer by early November. We are tracking all of it on the IBM i AI watch list for fall 2026, and this article gets revised the week of GA. The product profile lives in the directory at Power Autonomous Operations.

The hardware half of that same July 15 announcement, including local inference on entry Power 11 servers and Matrix Math Acceleration, belongs to our sibling sites: ai.as400system.com for AI on the platform and Power 11 for the systems themselves.