Entia non sunt multiplicanda praeter necessitatem.
Entities should not be multiplied beyond necessity.
‍
– Occam's razor

‍
If you watched the World Cup final, you will have noticed a departure from the usual 15 minute halftime break: a star-studded, Super Bowl-styled extravaganza featuring Madonna (with Ronaldinho and Ronaldo, to boot), the Muppets jamming to White Stripes’ Seven Nation Army, BTS, a dour Justin Bieber, Shakira and Burna Boy, ending with Mr. Coldplay himself, Chris Martin, and the PS 22 chorus singing about believing in love.

I do believe in love, but I was just hoping to watch some good football (wish not granted, unfortunately). The whole thing felt… too much.

I kind of feel the same way about how the industry is throwing AI at everything. Find the weak point in the process, and bolt another model or skill or agent onto it. Gartner calls this “agentic sprawl” — and it’s creating security issues and, naturally spawning an entirely new class of solutions. None of this solves the root cause, but it sure feels good to have done something, anything.

The maxim, “if you’re going to do something, do it well,” may have featured in your childhood — it certainly featured in mine. And if you (grudgingly, usually) take heed of it, you’ll learn that doing something well rarely means doing more of it. Most of the time it means stopping to figure out what the job really requires, perhaps going back to first principles, zero-based budgeting, MECE — pick your framework or technique of choice. And then teasing apart cause and effect. That's much harder, but so much more effective, even (especially?) when it means rethinking the order of an existing process.

Mainframe modernization has its own version of this, and it's a familiar one to anyone who's lived through a rewrite. Understand the system, extract the rules, build the new system, then verify: a testing phase, a UAT cycle, maybe a parallel run, scheduled for the end once there's something to test against. When that phase turns up mismatches, as it reliably does, the response is to pull in all those enticing new AI tools. Which in fact generates yet another testing cycle, followed by another set of discrepancies, followed by another manual sifting through the issues. It’s the same cycle, just fluffed up with new tools, incrementally faster for sure, but not solving the root problem.

The alternative reorders the process rather than adding to it: verifying that the behavior of the new code matches the behavior of the old system as it’s being written. Nothing ships without already being proven equivalent. Once you have that model in hand, then you can apply the bright, shiny new technologies and automate the heck out of it. And that’s exactly what we’ve done: applying harness engineering, aka dark software factory, techniques tuned to the specifics of mainframe modernization.

The software factory approach retains humans at two critical points: the seed, what you're building, and the harness, how you prove you got it right. Prove equivalence as the code is written, with the AI tooling running within a better approach, rather than added atop a sequence that’s proven to be sub-optimal. Less uncertainty; vastly more speed. Now that’s a beautiful game.

‍

News and Views

We’re not the only ones who think sprinkling AI fairy dust on current modernization approaches is doomed to failure. Gartner predicts (subscription required) more than 70% of mainframe exit projects started this year will fail, arguing that leaders are overestimating what generative AI alone can do and that vendors are bolting AI onto offerings regardless of whether it actually improves outcomes. That said, they’re conflating two issues: the problem of unverified generation versus whether modernization should happen at all. Baby and bath water.

When it comes to agentic sprawl, Gartner is predicting that, by 2028, “an average global Fortune 500 enterprise will have over 150,000 agents in use, up from less than 15 in 2025.” Naturally, they have some doubts as to whether agents can secure themselves (subscription required).

Speaking of less is more, Sam Learner leans into a similar argument in his FT Magazine piece, Who cleans up after the vibe-coding party? (subscription required). He follows cURL’s Daniel Stenberg, who shut down his bug bounty program this year, buried under AI-generated reports that took real human hours to check for every legitimate one. His take on the tools — that they’re better at finding problems than fixing them — is sober reading.

‍

From the Orchard

Imogen’s Summer 2026 Release, which shipped earlier this month, now receives output from AWS Transform, adding to our existing Google MAT integration. Business rule extraction (BRE) tells you what your system is designed to do; Imogen works from what it actually does, rewriting each component and running real production data through old and new code to prove they behave identically. Our VP of Product, Ingo Wiegand, details all that’s new in this latest drop here. We’ve also announced a partnership with Leidos: read the full press release here.
Our CTO Roberto Ostinelli describes how Imogen’s autonomous build-and-verify pipeline, aka Mechanical Orchard’s software factory, works. Once you’ve read about why modernization particularly benefits from a software factory, head over to Rachit Awasthi’s illuminating take on how dark factories can mean seeing even more clearly.

Our webinar last month, How AI agents actually make modernization work, featured Mechanical Orchard’s Edward Hieatt, David Yahalom from Google Cloud, and Michael Ljung from Thoughtworks, moderated by Arun Batchu. Full recording here. Arun also presided over an IRL version, over drinks and canapés at Nobu in Dallas two weeks ago, discussing the issues of trust: now that generating code is fast and cheap, “What will make you confident you can trust the new code generated by AI?”

‍

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

‍

—

View all newsletters here