An AI coding agent may need access to some of a company’s most valuable material: its source code, application structure and development tools. Where that agent runs can be as important as what it can generate.
On 1 October 2026, IBM announced a self-hosted deployment option for IBM Bob, its agentic software-development platform. The offering lets organisations operate the backend on infrastructure they manage, including on-premises and private-cloud environments. IBM’s release notes say general availability began on 24 September. IBM’s announcement; Release notes and availability date
“Agentic” means the software can use tools and work through tasks, rather than only suggest the next line of code. The announcement concerns deployment and governance, not a newly proven level of coding accuracy.
The backend moves, and the model choice still matters
The self-hosted backend runs on Red Hat OpenShift, a platform for managing containerised applications. Developers retain Bob’s existing client tools while the organisation manages the backend services and integrations. IBM’s deployment documentation
IBM describes two model routes. One connects Bob to a supported cloud model service through the organisation’s account. The other runs a supported open-weight model on its own GPUs. GPUs are processors commonly used for AI computation. Model deployment choices
This is an important data-boundary distinction. With the cloud-model route, model requests—including their code context—still leave the local backend for the chosen cloud service. Hosting Bob’s application backend locally does not by itself make every part of the workflow local. IBM’s description of request flow
Air-gapped operation requires a local model
An air-gapped environment is isolated from external network connections. IBM supports this configuration using a model on the organisation’s own GPUs. Its announcement also describes hybrid configurations that connect to external model services. The two arrangements have different operating boundaries. Local and hybrid model options
Open weights make a model’s trained parameters available for deployment under its licence. They do not automatically mean unrestricted reuse, free infrastructure or disclosure of every training detail.
Our assessment is that the meaningful change is architectural choice. Organisations handling sensitive code can consider an internal agent platform alongside their existing rules for model access, network connections and tool permissions.
Control comes with a new operations job
IBM’s documentation places responsibility for upgrades, scaling and availability on the self-hosting organisation. It also gives that organisation responsibility for configuring networking, storage and identity controls. Those tasks remain with IBM in the managed service. Responsibilities in each deployment model
The local backend includes identity services, an inference gateway, audit logging and usage metering. These features can support governance; they do not replace decisions about which repositories and actions an agent may access. Self-hosted backend components
The practical tests are familiar engineering questions: whether the selected model handles real workloads, whether generated changes pass review, and whether operating the platform is worth the additional effort. Self-hosting changes where the organisation exercises control. Demonstrating useful and reliable work remains a separate requirement.
Featured image: Representative software-development photograph; not a screenshot of IBM Bob. Photo: James Harrison / Unsplash.
Leave a Reply