Two reports from the same day show the same line of development from different directions: AI support in programming is increasingly moving into your own infrastructure. IBM is now also offering its assistant Bob, which helps with the modernisation of COBOL estates, as a self-hosted variant. At the same time, a field report describes how the open-source AI agent Hermes can be set up locally and what practically matters in the process.
For companies with sensitive data or long-established legacy systems, this is the decisive question. Not whether a model helps with refactoring, but where the source code flows in the process.
IBM Bob: modernisation without a detour via the cloud
Bob is aimed at organisations that operate applications based on COBOL – a programming language that has been used for decades, above all in banks, insurance companies and public administrations, and often carries business-critical processes there. This code is usually the most sensitive a company owns: it embodies business logic that is documented nowhere else.
IBM now offering Bob as a self-hosted option addresses this point directly. Companies can operate AI-supported software development in their own infrastructure instead of handing codebases over to an external service. For regulated industries, this is often less a question of convenience than a prerequisite for being allowed to evaluate such tools at all.
In terms of content, the task remains the same: understanding existing code, documenting it and gradually transforming it into maintainable structures. What is new is where it runs, not its purpose.
Hermes: what a local agent demands in practice
The second perspective comes from the open-source world. Hermes presents itself as a learning-capable AI agent, i.e. a system that not only provides text suggestions but independently calls tools and works through multi-step tasks. In a self-test, the agent was set up locally and then examined for what actually matters with models, tools and hardware.
These three keywords are the real message. A locally operated agent is not a product that you install and use, but a combination of decisions:
- Model choice: Which language model runs locally determines quality and response times – and there is no single option that suits all tasks.
- Tool integration: An agent is only as useful as the interfaces through which it is allowed to access repositories, files or build processes.
- Hardware: Local operation shifts the resource requirement into your own premises. Computing power is therefore no longer an abstract item on a cloud bill, but a concrete purchase.
The title of the field report plays ironically with the question of whether such an agent makes IT specialists redundant. The test setup itself points in a different direction: anyone who operates an agent locally needs more expertise, not less – just in a different place.
What distinguishes the two cases
Bob and Hermes do not solve the same problem and are aimed at different organisations. Bob is a vendor offering with a clear focus on legacy modernisation, Hermes an open-source, generalist agent that is being tried out in a self-test. One is a procurement decision, the other initially an experiment.
What both have in common is the mode of operation. In both cases, the code stays where it is created. This shifts the discussion from the question "May we use AI in the development process?" to "Under what conditions do we operate it ourselves?".
No reliable information on performance data, licensing models or specific hardware requirements is available from the reports at hand. Anyone seriously evaluating either option should clarify these points early in the evaluation process – in particular the actual resource requirement in continuous operation.
Assessment for project planning
Self-hosted AI tools change the cost structure. Instead of usage-based fees, there are investments in infrastructure and operation, plus ongoing effort for updates, model maintenance and access rights. This can pay off, but it does not have to – the calculation depends heavily on usage intensity.
A second point concerns expectations. Agents that operate tools independently require clear boundaries: which systems may they read, which may they modify, and who checks the result? These questions belong before the introduction, not after.
For web and software projects, this means concretely: the data protection reservation loses weight as a blanket counter-argument when AI assistance can also run in your own data centre. In its place comes a sober operational question – hardware, model maintenance, access concept – that can be planned like any other infrastructure decision. Anyone who has to replace or document older systems thus has one more approach to choose from, but should test it on a manageable sub-area before it becomes the basis of an entire modernisation project.