Every enterprise I talk to has a version of the same problem. The systems that run the business, the billing platforms, the provisioning engines, the ordering flows, are decades old, and the people who truly understand them are a few years from retirement. The succession plan, where one exists, is a wiki nobody updates and a hope that somebody younger volunteers to learn COBOL.
Here is what most of those companies overlook. The knowledge is not only in the veterans' heads. A large share of it is sitting in version control, sometimes in repositories as old as SourceSafe, in the form of every change ever committed to that system. Thirty years of diffs, commit messages, and bug fixes is a written record of how the application actually works and why it works that way. No human will ever read it all. An agent will, happily.
This changes the knowledge capture problem completely. Instead of asking a retiring engineer to spend their final year doing brain dumps, you point an agent at the full history and let it reconstruct the story of the system, module by module, decision by decision. It will surface things the original authors do not remember writing. Then you bring the veterans in as reviewers. Correcting a draft is a far better use of a scarce expert's time than producing one from nothing, and in my experience the people who hold this knowledge are not hoarding it. They have simply never had anyone capable of absorbing it.
The same pattern applies outside engineering. The field technician who has seen every failure mode, the operations person who works the exception queue every morning, they all run on accumulated context that lives nowhere durable. Capturing it is the hardest and most valuable work in enterprise AI, which is exactly why organizational context remains the unsolved problem, and why the companies that capture it will find that context, not model access, is the moat.
The retirements are already scheduled. The history is already written down. The only question is whether you mine it while the reviewers are still in the building.
Key takeaways
- Every enterprise has critical legacy systems understood by a handful of veterans approaching retirement, and no succession plan that actually captures what they know.
- Decades of version control history contain a written record of every change ever made, including context the original authors no longer remember.
- Agents can mine that history into a living knowledge base, then let the veterans correct and enrich it while they are still around to do so.
FAQ
How can an agent extract knowledge from old version control systems?
An agent can walk the entire commit history of an application, read every diff and every commit message, and reconstruct a narrative of how the system evolved and why. Even sparse commit messages carry signal when you can analyze thousands of them alongside the code changes themselves.
Does this replace talking to the veteran engineers?
No, it changes what those conversations are for. Instead of asking veterans to do a brain dump from scratch, you show them a reconstructed history and ask them to correct it and fill gaps. Review is a far better use of their remaining time than recall.
Sources
Related Essays
Organizational Context Is the Hardest Problem Nobody Has Solved
Context management is the layer most consistently built in-house and least well served by vendors. It is not a search problem. It is a knowledge management problem.
Context Is the Moat — Don't Give It Away
If you upload your entire business to a frontier model provider, you have handed them the playbook. Keep context local. Use frontier models as the engine.
The Real Product Is the Graph, Not the Interface
When your polished front-end is a byproduct of an underlying knowledge graph, be honest about which layer is the product and which is the tech demo.
Drowning in pull requests that need your review? Try Tembo Review, a beautiful AI-assisted PR review tool unlike anything you’ve used.