Make "Keep the Code In-House" the Default — IBM Bob Pulls the Entire Agentic Dev Stack Behind the On-Prem, Air-Gapped Wall
IBM has begun offering its coding-agent platform "Bob" as a self-hosted deployment that runs in on-premises, sovereign, and air-gapped environments. We break down this move — along with its prerequisites and open questions — that hands regulated industries unable to let code leave the building the o
Bringing the Agent to Code That Can't Leave the Building
On October 1, 2026, IBM announced that IBM Bob, its agentic software development platform, can now run as a "self-hosted" deployment inside environments the enterprise manages itself. Supported setups span on-premises, private cloud, sovereign cloud, and even air-gapped environments cut off from any external network. The intent is clear: for banks, governments, and regulated industries that have hesitated to send source code or regulated data to a third-party AI cloud, IBM is offering a way to run AI-driven development while keeping the code in-house.
Bob itself debuted in 2025 as a multi-agent development platform that goes well beyond code generation. On top of planning, implementing, testing, and deploying changes, it assigns agents across the entire software development lifecycle (SDLC) — including modernizing legacy assets accumulated in Java and on IBM i / IBM Z. What's new about this announcement is that the whole stack can now be completed "behind your own walls."
Where It Runs, and Which Models You Can Use
The models available to you depend on how you deploy it. Here's how the publicly available information breaks down.
| Deployment type | Target environment | Main available models |
|---|---|---|
| Self-hosted | On-premises / private / sovereign / air-gapped | NVIDIA Nemotron, Poolside Laguna |
| Hybrid | Offload only low-sensitivity work to approved external services | Claude Sonnet 5.0 / Claude Opus 4.8 / Gemini 3.7 Flash / OpenAI GPT 5.6 Sol |
A hybrid configuration lets you switch routing on a per-workload basis. Sensitive code is handled by models running locally, while less-constrained work is routed to approved external services. The idea is that developers keep a consistent Bob experience, while the security team retains control of the architecture running underneath it.
Shifting the "Build vs. Buy" Calculus for Regulated Industries
This announcement lands squarely with the companies that have so far held back from adopting AI coding. Many AI development tools have assumed that code and context get sent to an external service. For companies that cannot let their intellectual property or customer data leave the building, that was a non-starter from the outset.
IBM's own research reportedly found that 68% of executives say it is difficult to meet region-specific data residency and sovereignty requirements. The self-hosted option is a move to tackle that "can't send it out" barrier head-on. Neel Sundaresan, General Manager for AI and automation, stated that "organizations need AI that runs inside an environment they control — especially when handling sensitive code and regulated data."
Expanding coding-agent adoption beyond the single option of "buying an external SaaS" to a form that is "kept under control and operated in-house" means regulated industries gain one more realistic option for their digital transformation.
Caveats to Check Before Adopting
That said, self-hosting isn't a matter of "drop it in and it runs." You need to keep in mind the prerequisites that coverage has pointed out.
- Models used in self-hosted deployments must be procured, licensed, and hosted by the adopting company itself. Responsibility for operating the models shifts to your side.
- Pricing is not public, and IBM directs prospects to individual demos and inquiries. The cost structure is hard to estimate in advance.
- Some capabilities, such as routing across multiple models, are described as "planned on the roadmap," so it's wise to evaluate shipped features separately from planned ones.
The Issues Behind the Convenience
"Running behind your own walls" is reassuring from a security standpoint, but it also shifts responsibility toward the enterprise. You take on operating, updating, and monitoring the models yourself — tasks you had left to an external cloud — and even in an air-gapped environment, reviewing and managing permissions for the code agents generate and run still remains. "Code doesn't leave the building" and "the agent behaves safely" are separate problems, and it's worth keeping a clear eye on the fact that the former doesn't eliminate the need to prepare for the latter. In adopting it, you need a posture that weighs not just the marketed sovereignty but also how much of the operational burden your own organization can actually take on.
References: IBM Newsroom / InfoWorld / SiliconANGLE / MarkTechPost
