“Acting without knowing takes you right off the cliff.”
— Ray Bradbury

‍

‍

Back in 1973, two UC Berkeley professors wrote a paper called Dilemmas in a General Theory of Planning, arguing that social problems can’t be solved with one clear, scientific answer: no “silver bullet” as it were. The authors, Horst Rittel and Melvin Webber, fault the success of the scientific method in solving bounded problems—say, vaccines to suppress measles, or rockets that send humans to the moon—as creating a mistaken belief that social and political challenges could be tackled the same way.

‍

Weirdly enough, mainframe modernization exemplifies what Rittel and Webber called a wicked problem: it resists precise definition, has no clear end point, and every attempted solution reshapes the challenge itself. Each organization’s legacy environment is unique, and modernization decisions—whether to rewrite, refactor, or replace—are shaped by competing priorities, cultural norms, and risk tolerance.

‍

If you’ve been reading this newsletter, you know we view mainframe modernization as an anthropological endeavor as much as a technical one. The confluence and accretion of multiple systems over the years has created technology stacks that resemble social constructs more than “clean” technological constructs. Put another way, mainframe modernization is social science masquerading as computer science.

‍

The prevailing model—system integrators orchestrating a suite of tools—starts from the wrong question. It assumes the problem is “how do we convert code faster?” But modernization’s core challenge isn’t throughput; it’s framing. If you define the problem as code translation, you’ll measure lines converted, tests generated, and environments spun up—and still fail.

‍

Applying GenAI to this old framing can actually make things worse: we’ve now created a new class of tools to solve problems that didn’t even exist before. Rules-engine tools, designed to rein in the non-determinism of large language models, are solving for the side effects of using LLMs in the first place. It’s a perfect illustration of how wicked problems multiply when treated as technical puzzles: every “fix” generates new complexity. And before you know it, you’ll be beholden to a single vendor to help juggle a vast array of software tools.

‍

An alternative framing is to recognize that mainframe modernization is an ongoing negotiation between systems, people, and priorities. Know that the ultimate outcome is a malleable, easily-evolvable structure. In part that means, instead of adding more technology, frameworks, and abstractions, start with what’s working today. (We tend to neglect subtraction as a way to change things because of how our brains are wired.) Let the system’s running behavior define the specification and the tests. And chip away at that big block of uncertainty with smaller doses of actual workloads in production.

‍

Wicked problems don’t have to be scary—only when we mistake them for simple ones do they turn into monsters.

‍

News and Views

Black swan events can create wonderful natural experiments. When the pandemic hit, states were charged with administering federal unemployment insurance funds. The Federal Reserve Bank of Atlanta took the opportunity to compare states with COBOL systems and those without, estimating that the 28 states with COBOL systems reduced US GDP by $40 billion. This would make modernization (cost: roughly $500 million) a screaming deal.

‍

After living in the UK for a few years, I find the ubiquity of cash transactions in the U.S. quaint. But now I understand: the FT’s Brendan Greeley writes that last time U.S. banks were compelled to invest heavily in compliance, the effort was so resource-intensive, it was simpler to fund lobbying efforts to keep things the same.

‍

From the Orchard

In his latest blog, our CTO, Roberto Ostinelli, posed a thought experiment: in a world where LLMs and AI-assisted code generation are taking on larger roles, what does idiomatic code even mean? And, taking it a step further, if idiomatic really means “maintainable,” for whom is this maintainable? These questions highlight the need for rigorous testing and verification—and reframe the exam question.

‍

We talk a lot about modernization as a competence, and are particularly keen on the concept of “continuous modernization” that Gartner coined in 2018. And while Imogen is a platform that spans behavioral discovery, rewrite, and equivalence testing, it’s still nice to be recognized in one of their latest roundups of tools for AI-assisted modernization.

‍

Speaking of talking a lot, our COO Edward Hieatt is joining Leda Glyptis and execs from Thoughtworks and Google for a webcast on making the business case for modernization. It’s happening on November 13; register here.

‍

And if you want to participate in the conversation directly, we’ve joined forces with our partner Thoughtworks to kick off a networking and event series called Unfinished Business. Coming to a city near you in 2026, these gatherings offer space for frank conversations about the challenges of legacy modernization and explore ways to get things unstuck. To get the schedule and an invitation, drop me a line.

‍

Curious to learn more? Say hello@mechanical-orchard.com.

‍

—

*Issue first published on Octber 30th, 2025. View all newsletters here

‍