Thirty or forty years of RPG and Db2 isn't legacy code. It's how your business actually runs: your pricing, your credit rules, your exceptions, the decisions nobody ever wrote down anywhere else. You don't ship that to someone else's endpoint to be read. So when the first wave of AI coding tools showed up asking you to do exactly that, most IBM i shops made the right call and waited.
On October 2, IBM took that reason off the table for Bob, its AI platform for building, modernizing and maintaining software. Bob can now run inside your own infrastructure: on-prem, in private or sovereign clouds, or fully air-gapped. You can run supported models on your own hardware, or set up a hybrid that calls out to supported external model services when you choose to.
Where the AI runs matters as much as what it can do.John Duigenan, Distinguished Engineer, IBM, to IBM Think1
That line matters more than any feature in the announcement. When the largest vendor in our world says it out loud, placement stops being an argument you have to win inside your own company. It becomes a product option on the order form.
What actually changedThree things, in plain terms.
- The code can stay home. Self-hosted Bob keeps your source and your data inside your own environment. The tool comes to the code instead of the code going to the tool.
- You decide where the thinking happens. Run models on your own infrastructure, or run a hybrid that uses external services for some work. That isn't one yes-or-no decision anymore. It's a decision for each kind of work.
- The category opens up. IBM says self-hosting turns Bob from a tool many sensitive teams couldn't even consider into one they can adopt. For regulated IBM i shops, the question stops being whether AI can touch your code. It becomes what you allow it to do once it's inside.
This isn't a niche worry. IBM's own research puts a number on it:
Sovereignty used to be the buyer's objection. Now the seller is putting it in the brochure.
What it doesn't changeThe parts that are still yours to do.
Readiness doesn't come in the box.
Bob being able to sit beside your RPG doesn't mean your estate is ready for an agent to work in it. You still need to know where the business rules live, who understands them, and what can be changed safely.
"Hybrid" means you need a policy.
If some work can call outside models, someone has to decide which work, and someone has to approve that decision. Write the policy before the pilot, not after the first surprise.
An agent inside the box is still a new identity on the box.
Letting Bob in is an access decision. And access is the part of this story IBM i is better equipped for than almost any platform out there.
The IBM i advantageYou already know how to govern this.
The risk the whole industry is now scrambling to manage is inherited authority: an agent that quietly acts with the full permissions of whoever launched it. Most platforms are buying new tools just to see that, let alone limit it.
IBM i has managed this problem for decades, and IBM's own Bob documentation says so plainly.3 Bob reaches your system through the user profile of its Code for IBM i connection, over SSH. Anything that profile is permitted to do, Bob can do too, and IBM's guidance is to give that profile the least privilege necessary. In other words, Bob's authority on the box is exactly the profile you hand it. The controls are already native:
Give the agent its own user profile.
Not a developer's profile, and never one with *ALLOBJ. A dedicated, least-privilege profile for Bob is the simplest and strongest move you can make. One practical note: the Bob connection doesn't support single sign-on today,4 so plan how that profile's credentials or SSH keys are issued and protected.
Watch adopted authority.
Programs that adopt their owner's authority are a quiet way for an agent to reach things nobody meant to grant. Before Bob calls into your applications, know which programs adopt and what they can reach.
Let the audit journal answer "who did what."
QAUDJRN records actions at the operating-system level, beneath the application, where nothing can talk its way around it. Other platforms have to build that record; you already have it. IBM also notes Bob marks its activity with a value you can use to track and audit what originated from Bob.3 With a dedicated profile, that's a clean answer to "what did the agent do?"
Lean on journaling and commitment control.
Changes an agent proposes can be made traceable and reversible at the database, not just reviewed in a ticket.
The AI suggests. A person decides.
That principle runs through everything we do. Self-hosted Bob makes it practical on the most sensitive code you own, because the platform it runs beside already enforces the rules.
What to do MondayThree moves before the first prompt.
Revisit your AI-coding policy.
If it still reads "no external tools on source," it was written for a choice that no longer exists. Rewrite it around placement: what runs in-house, what may call out, and who approves the difference.
Pick the agent's identity before its first task.
Create the profile, scope its authority, and confirm the audit journal is capturing what you'll want to see.
Start with one contained piece of work.
A well-understood program, a documentation pass, a test-coverage gap. Learn how Bob behaves on your code before it touches anything that moves money.
Want to know how ready your estate is for an agent you've decided to let in?
Take the IBM i readiness read