IBM Puts Its Bob Coding Agent Behind the Firewall: Self-Hosted and Air-Gapped Deployments Are Now Generally Available

IBM has made a self-hosted deployment of Bob, its agentic software development platform, generally available, giving regulated enterprises a way to run an autonomous coding agent entirely inside their own perimeter — on premises, in a private or sovereign cloud, or in a fully air-gapped network with no external connection at all.

Bob is not an autocomplete plug-in. It covers the full software lifecycle: reading and mapping an existing codebase, planning work, decomposing it into subtasks, spawning subagents, executing and testing changes, and handling legacy modernization. The GA package ships the full stack — Bob backend, identity management, an inference gateway, audit logging and usage metering — running on customer-managed infrastructure including Red Hat OpenShift clusters, alongside the IDE experience, the BobShell command-line interface, parallel tool calling, and the agent harness with its skills and modes.

The model layer is bring-your-own-license. For fully air-gapped deployments, where code and context must never leave the building, IBM currently supports NVIDIA's Nemotron and Poolside's Laguna running on customer GPUs. Hybrid configurations keep the Bob backend on-premises but route selected workloads through private connectivity to approved external services, including Claude, Gemini and OpenAI models, chosen per workload based on sensitivity.

IBM's internal numbers explain the pitch. Bob has run inside IBM since June 2025, growing from 100 developers to more than 80,000 users, with surveyed employees reporting an average 45 percent productivity gain. Those are vendor-reported, self-surveyed figures from IBM's own workforce rather than an independent benchmark, and should be read accordingly.

The commercial logic is aimed squarely at buyers who have spent two years watching competitors ship AI-assisted code while their own compliance teams said no. IBM's own survey work cites 68 percent of executives naming data residency requirements across geographies as a major challenge, and Gartner has predicted that 35 percent of countries will be locked into region-specific AI platforms by 2027. GitHub's Copilot coding agent, by contrast, runs only in GitHub's cloud and does not support GitHub Enterprise Server deployments.

The timing is not accidental. Between July and September, autonomous coding agents were implicated in a database deletion at PocketOS and an infrastructure breach at Hugging Face — incidents in which damage occurred without human input. IBM's answer is to make governance part of the deployment rather than an afterthought: BobShell writes self-documenting audit trails in real time, and the workflow includes prompt normalization, sensitive-data scanning, real-time policy enforcement and AI red-teaming.

Optional Premium Packages extend Bob to Java modernization, IBM i and IBM Z mainframe estates — precisely the long-lived codebases that enterprises most need to modernize and least want to expose to an external API. IBM has not published pricing; the offering routes through enterprise sales.

The honest trade-off is that sovereignty is not free. Customers sourcing and hosting their own models take on GPU capacity planning, model lifecycle management, patching, scaling and on-call responsibility for the platform itself. IBM's bet is that for banks, defense contractors and public-sector bodies, that operational bill is now cheaper than the alternative: sitting out agentic development entirely.