Let a Million Flowers Bloom

The glorious, messy, occasionally unhinged experimentation happening in software development right now

I was lurking in a team chat the other day, watching a group of senior engineers debate the finer points of how to give AI agents better memory. One camp was using a structured task-tracking database, kept local to their machine, never committed to git. Another had rigged up a MEMORY.md file with custom hooks that auto-indexed what the AI had learned, capped at exactly 200 lines so it would auto-load. A third was tying ephemeral AI work plans to git worktrees through some clever .git directory linkage that the AI itself had suggested.

These are not junior developers fumbling around. These are experienced engineers with decades of collective practice, and they are all doing something fundamental to the highest traditions of engineering: they are tinkering. Not tinkering with their code — tinkering with how they write code. Tinkering with the relationship between human intention and machine execution. And every single one of them has a slightly different rig.

We are at the “let a million flowers bloom” stage, and I’m hear to tell you the garden is wild.

The Experimenters

What struck me about this particular conversation wasn’t the technical details — those are interesting but contingent, likely to be obsolete in six months when the underlying models shift again. What struck me was the spirit of the thing. Here were working professionals, shipping real software at a real company, and they were swapping notes on their personal experimental setups with the enthusiasm of ham radio operators comparing antenna designs in 1955.

“Would you be willing to let me look over your shoulder next time you start a new session?”

“Yes I can record and send it here.”

This is not how software engineers typically talk about their tooling. We are, as a profession, notorious for picking a tool and then defending it like a theological position. Vi versus Emacs. Tabs versus spaces. IntelliJ IDEA versus a whole bunch of extremely wrong people. The fact that these conversations now sound less like religious warfare and more like field notes from a shared expedition tells you something important about the moment we’re in.

Nobody knows what works yet. And everybody knows that nobody knows. And I am proud of my profession when I report that has made us curious rather than defensive.

The Taxonomy of Tinkering

If you zoom out, you can start to see rough categories of experimentation, even if no two setups are alike.

There are the memory architects — people obsessed with the question of how to give AI agents persistent context across sessions. They’re building custom databases, maintaining hand-crafted markdown files, wiring up hooks and indexing schemes. The core problem they’re solving is that every AI conversation starts with amnesia, and they refuse to accept that.

There are the workflow sculptors — people who are redesigning their development loops around what AI can and can’t do. They’ve figured out that the interesting question isn’t “can the AI write this code?” but “what’s the right unit of work to hand it?” One engineer I know uses AI for anything under a story point and manually tracks everything above that in a structured hierarchy. The boundary is fuzzy and he admits the AI misjudges the threshold about 65% of the time, but he’s iterating. He’ll get there.

There are the context engineers — people who have realized that the bottleneck isn’t the AI’s intelligence but its awareness. They’re spending serious energy figuring out what the AI needs to know, when it needs to know it, and how to keep that knowledge fresh without blowing out token budgets. This is, quietly, becoming a genuine subdiscipline.

And then there are people like me, who occasionally wander off the map entirely.

My Weirdest Experiment

My personal strangest experiment so far has not been in my usual wheelhouses of programming or analytical thought at all. It was testing some hypotheses about human/LLM centaur capabilities by seeing how well Claude could help me write a comic erotic novel.

Yes, really.

The reasoning was actually sound, I promise. I was interested in what happens when AI collaboration can’t collapse into “just have the AI do it” — which is what happens with a depressing number of so-called AI-assisted projects. Comedy is notoriously difficult for language models; they can maybe identify what’s funny but can’t reliably generate it, usually producing something at the level of “guy at a bar who thinks he’s amusing after a couple of Heinekens”. And the, shall we say, spicier content is something the models are specifically disinclined to produce. Once the text had to get down to the “Tab A into Slot B” level, Claude would reliable tell me I was on my own. So the only way the project could work was as a genuine collaboration where the human brought irreplaceable capabilities to the table. It was also intended to be intentionally difficult, as sex comedies often fall flat. In the words of the sage ”Sex is funny. Comedy is funny. The French are funny. But there are no funny French sex comedies.”

The results were interesting-but-not-great, but that was back in July, which in AI time is roughly the Mesozoic Era. I keep meaning to resurrect it and see how the newer reasoning models do. (For the record, only personal tokens from my Max plan were harmed in the making of that experiment, not any tokens from The Day Job.)

The point isn’t the novel: “My Best Friend’s Bachelor Party”, an homage to the 1997 Julia Roberts looks-like-a-rom-com-but-isn’t vehicle “My Best Friend’s Wedding”, only with a lot more boning. The point is that the instinct to try it — to probe the edges of a new capability by deliberately seeking out the cases where it’s weakest — is the same instinct driving the memory architects and workflow sculptors and context engineers. We’re all doing the same thing: tapping along the walls of the possible, looking for where it gives way and where it holds.

Why This Matters

I’ve been writing software for thirty-five years, and I have never seen this much simultaneous experimentation happening at the practitioner level. Not during the rise of open source, not during the Agile revolution, not during the shift to cloud. Those were large, organized movements with manifestos and conferences and certifications. What’s happening now is more like what happens when a new continent is discovered and everybody grabs a boat.

Some of these boats are going to sink. Some of these experimental setups are going to turn out to be dead ends. The engineer maintaining a 200-line auto-indexed memory file may discover that next month’s model update makes the whole scheme unnecessary. The structured task database may get superseded by native capabilities. The worktree-linked epics may prove too brittle for real-world use.

That’s fine. That’s the point. The underlying substrate is changing fast enough that everything is contingent. The correct response to a landscape that shifts under your feet every few months is not to plant a flag and defend it — it’s to keep moving, keep probing, keep sharing notes with the other explorers.

What I find most admirable about my profession right now is precisely this willingness to look a little foolish. These are people who have spent careers building expertise, and they’re cheerfully dismantling their own workflows to see what happens. “I have had this happen about 65% of the time” is not something an insecure person says in public. That’s someone who has made peace with the fact that they’re running an experiment, not operating a production system.

The Garden

There’s a Taoist concept — wu wei, often translated as “effortless action” — that I think applies here, although probably not in the way any Taoist scholar would endorse. The idea is that sometimes the right thing to do is to stop trying to control the outcome and instead create the conditions for the right outcome to emerge.

That’s what a million flowers blooming looks like. Nobody is in charge of this. No vendor is dictating the One True Workflow. No methodology consultant is selling certifications. Engineers are just... trying stuff. Sharing what works. Admitting what doesn’t. Moving on.

It’s messy and inefficient and beautiful, and I think when we look back on this period in five or ten years, we’ll recognize it as one of the most creatively fertile moments in the history of the profession. Not because of any single breakthrough, but because an entire discipline collectively decided to become beginners again. Just starting. Every day, just starting.

That takes a kind of courage that doesn’t get celebrated enough. So here’s to the tinkerers, the experimenters, the people maintaining weird markdown files and local databases and git worktree hacks. Here’s to the ones writing comic erotica to probe the boundaries of human-AI collaboration. Here’s to everyone who looked at this strange new technology and said the only thing worth saying:

“Wanna see something cool?”

The garden is wild. The flowers are weird. And I wouldn’t have it any other way.

---

This post was constructed with the able assistance of Claude Opus 4.6, whose comedic genius remains as yet undiscovered.