What Flowering
forgot about flowers

The problem

The original Flowering pattern is a methodology for staged AI work. The shape: seed, then shape, buds, bloom, cross-cutting refinement, and finally — at the end — pruning. It explicitly named "endless bloom" as an anti-pattern. It explicitly warned that artifacts must not bloom forever. The structure was, in every respect, growth-shaped: each loop added to what came before, and pruning was a terminal cleanup phase.

It met its first stress test as the basis for a real consequential decision. A real-estate development team had circulated a 30-page requirements doc for a custom AI dashboard — covering pay application validation, draw package assembly, AIA G702/G703 lien-waiver matching, leasing pipelines, loan compliance, JV waterfalls, integrations to four production systems. Target ship: two months. Decision needed by Monday: build it custom, buy mature platforms with a thin layer on top, or pure-buy.

The agent applying Flowering produced a research project, three candidates, twelve testable claims, seven findings with content-hashed sources, and a one-page memo. Every artifact looked finished. None of them had been pruned.

The agent had blooms and only blooms. Asked to plan, it had already produced a four-stream framework (initiative, research project, Docket tasks, Governor ADR) before any work was done. Asked to estimate, it imported a human-consultant timeline of three working days for a thirty-minute job. Asked to draft a recommendation, it skipped seed and shape and went straight to a five-step polished plan — what the methodology itself names finality jump.

The agent was doing exactly what Flowering's anti-patterns warned against. The warnings did not prevent the behavior.

That is the diagnostic. Naming an anti-pattern is not a mechanism for avoiding it. The methodology had no continuous pruning verb — only a terminal one — and so the pruning never happened, because the work shipped before reaching the terminal phase.

The approach

The fix did not come from refining the warnings. It came from studying the biological process the methodology had named itself after.

Real flowering — the plant biology — is mostly the opposite of what the methodology described. A flower is biology's most concentrated example of overproduction-and-triage. An apple tree drops roughly 90% of its flowers as "June drop" before they ever ripen. The plant runs continuous source-sink accounting on every leaf and every developing fruit, abandoning the ones that go negative regardless of how much energy went into them. Apical dominance — auxin from the growing shoot tip — actively suppresses lateral buds so one main shoot extends and the rest stay dormant. Without that suppression, plants become bushes, not trees. Abscission — the controlled drop of leaves and flowers via the abscission zone — is hormonal and continuous, not janitorial. A tree that cannot abscise dies in winter.

None of this fits the growth-shaped methodology. Flowering v1 had pruning as a final phase. Biology has pruning as the other half of every metabolic gradient — not after growth, but during it, in every cell.

That observation forced two refinements, in sequence.

v2 — Port the biological mechanisms

The first attempt added eight biological primitives directly into Flowering's loop structure: a determinacy declaration upfront (annual or perennial — cymose vs. racemose inflorescences); an abscission pass at the start of every loop (drop before adding); an apical declaration per loop (name the one shoot extending, suppress the rest); a source-sink audit (drop what stopped earning); a quorum gate between loops (don't advance on a single signal — borrowed from bacterial quorum sensing); a commit-and-suppress phase (close the question once decided — borrowed from honeybee swarm decision-making, where dissenting scouts are vibrationally head-butted into silence after quorum forms); stigmergic state (work lives in shared environment, not in any agent — borrowed from ants and termites); and built-in dormancy phases (borrowed from sleep and glymphatic clearance).

v2 was an improvement. It was also brittle: the biological metaphors carried the argument. A reader who didn't trust biology had nothing to verify against.

v3 — Abstract to the underlying principle

The deeper move was to recognize that the eight biological mechanisms are not biology being imported into agent design. They are eight implementations of the same structural principle, which biology happens to instantiate well. The principle generalizes beyond biology — it shows up in machine learning (regularization, dropout, weight pruning), editorial work (kill your darlings), investment management (cut losers, run winners), forest management (prescribed burns), memory systems (garbage collection), cumulative knowledge (citation reweighting, theory abandonment), and software (maintainer-driven code pruning).

Naming the principle:

More compactly: growth without proportional and targeted pruning is decay disguised as productivity.

The evidence

Three things make CPS more than a slogan: the cross-domain pattern, the stress-test against counterexamples, and the refinement from named critics.

The cross-domain pattern

If the same principle shows up across this many independent domains, it is a structural law, not a metaphor:

Counterexamples that refined the principle

Six counterexamples were considered before publishing. Most refined CPS rather than refuting it:

Refinement from named critics

Two scaling-theory perspectives forced sharper claims.

Geoffrey West (scaling laws): organisms scale sublinearly — bigger animals are more energy-efficient per gram (Kleiber's law). Cities scale superlinearly — bigger cities are more productive per capita. Companies scale sublinearly like organisms. Where do current LLM agents fall? Anti-scaling. Context windows are hard caps; accumulated context adds noise; multi-turn performance degrades. Agents do not gain efficiency from accumulated structure the way organisms or cities do. Pressure on internal pruning is therefore inversely proportional to the system's scaling exponent. Agents need it most.

Stuart Kauffman (edge of chaos): productive complexity peaks at intermediate connectivity. Too sparse: stagnation. Too dense: chaos. CPS as initially stated had a floor (destruction ≥ unproductive growth) but no ceiling. Refined: pruning must be sufficient and targeted. Random pruning destroys productive structure as readily as unproductive. For agents, abscission is not "delete 30% of context per loop" — it is "delete the parts not being used in current attention."

The eight implementations

The biological mechanisms now sit as one valid implementation set for CPS. Domain-appropriate alternatives can substitute (an editor uses kill-your-darlings; a portfolio manager uses rebalancing; an ML engineer uses weight pruning). The names are mnemonic; the disciplines are load-bearing.

The trade-off curve

Growth-shaped methodologies were appropriate for a long time. They are not wrong; they are wrong now, for agents specifically, and the reason is operational rather than philosophical.

The reader knows they are on the new side when their agent's output exceeds the bandwidth of any external pruner.

The result

The methodology was rewritten in place. It now claims a structural law (CPS) instead of a set of biological metaphors, and the claim is testable: if a domain produces sustained value without proportional and targeted pruning, CPS is wrong. The original framing's central anti-pattern ("endless bloom") was a warning. The rewritten methodology's central primitive (abscission as the first action of every loop) is a mechanism. The difference is whether the discipline is enforced or merely described.

The published methodology also ships with a buildable prescription: bonsai, an external CLI whose only job is to challenge the calling agent to prune its working set systematically — asking source-sink questions on each retained item, returning a drop list with reasons, and writing abscission decisions to a persistent substrate so the next invocation sees the history. The CLI is itself an instance of the principle it enforces: stigmergic state (lives in the environment) and targeted pruning (asks ROI questions, not "what's old"). Build it once; every agent that talks to it inherits the discipline.

The real-estate-platform decision that triggered the case is now a worked example for the rewritten method — a perennial research project, with declared determinacy, recorded findings, content-hashed sources, locked criteria pending peer review, and an explicit ship gate at quorum. The agent that originally produced 12 claims for it would, under the rewritten discipline, drop to 6 in the abscission pass before claim 7 was even added.

How to know if it stuck

Three observable signals would tell us CPS is doing the work it claims:

1. Artifact size shrinks before it grows. A method-shaped session produces smaller, sharper output earlier in the loop, not larger output that gets cut at the end.
2. Closed questions stay closed. Once a quorum gate has fired and commit-and-suppress is in effect, the same debate does not reopen on second thought.
3. Cross-domain practitioners can verify the principle in their own field. An ML engineer recognizes regularization as CPS in their domain. An editor recognizes line edits. A fund manager recognizes portfolio rebalancing. None of them have to take biology on faith.

What is not yet validated

The case study would be incomplete without naming what the methodology has not earned:

Per the methodology's own discipline: quorum has not been reached. The published version ships as a draft, not a stable contract. The next signals are real-world use across multiple practitioners and review by someone with genuine standing on scaling laws or complexity theory.

The prediction is that methodologies which include destruction as a first-class primitive will displace ones that don't, the same way typed modular code displaced single-file scripts as project size scaled. The forcing function is agent throughput. The mechanism is CPS. The test is the next perennial.