Cut into the IT infrastructure of any large institution and you'll find something like a geological cross-section. At the bottom, a mainframe, still running, still processing transactions. Above it, a relational database from the nineties. Above that, a web portal from the early 2000s. At the surface, a modern API facade exposing clean endpoints to mobile apps and partner integrations. Each layer was introduced to replace the one beneath it. None of them did.
The IRS's Individual Master File has been running since 1961. Replacement estimates started at $549 million and climbed past $1.25 billion as the scope of what the system actually does became clearer. The GAO found that data from one system must be re-entered into another because the applications can't interface directly. People perform that re-entry. Call it clerical labor and you'd miss most of what's happening: humans bridging systems that were never designed to talk to each other, carrying context about which fields matter, which errors can be ignored, which sequences produce correct results.
As layers accumulate, the logic migrates in a particular direction. A rule that was once in documentation becomes a line of COBOL. A line of COBOL becomes a stored procedure. A stored procedure becomes a manual override performed by someone who's been there long enough to know the system produces the wrong output on the third Tuesday of the quarter unless you clear a specific field first. The adaptation looks like coping. It looks like habit. So it never gets documented, and the next system can't model what was never made explicit. The UK's NHS spent over £10 billion on a unified replacement before abandoning it in 2011, largely because the new system couldn't accommodate how clinical work actually happened in each setting. Peer-reviewed studies of electronic health records found an entire category of workaround labeled "no correct path," where the right option simply doesn't exist in the interface. People invent one. Over time, the invention becomes the process.
Enterprise architecture has a formal name for this incremental layering: the strangler pattern, after a fig vine growing around a host tree. The vine was always supposed to be transitional. The tree rarely dies. Agents are the newest vine, and the expectation is familiar: this layer finally breaks the chain. A University of Washington analysis of agents operating on state agency workflows describes something less dramatic: agents training on "fields that reject data until another field is saved, warnings that must be dismissed before real progress begins, and confirmation steps that look identical but encode different logic." The agents learn the workarounds. They have to. The workarounds, after all, are the process.
What lived in a person's memory now lives in an agent's learned behavior. The dependency has changed form. It is still a translation layer, and what it translates is still the accumulated logic of every layer beneath it, including the human one. Whether encoding workarounds in agent behavior makes them easier to see or harder, more brittle or more durable, won't resolve quickly. Each previous layer took a decade or more to reveal what it had actually absorbed.
- Agents absorbing institutional memory: Amazon's AGI Lab is training agents on high-fidelity simulations of legacy systems, learning the quirks and silent dependencies that only human operators used to carry.
- California's persistent workarounds: Despite a major Deloitte modernization contract, California EDD audits found employees continued using workarounds on patchy old systems that the new layer was supposed to replace.
- Hallucination meets legacy fragility: Practitioners are documenting how agents that encounter a failed API call will hallucinate data rather than report the error, a failure mode that compounds when the underlying system already produces unreliable outputs.
- Translation is permanent, not transitional: Enterprise integration architects are observing that from COBOL copybooks to XML schema transformations to JSON across REST APIs, the core translation problem persists unchanged even as the surrounding technology evolves.