Count the cascade, not the keystone.
In 1963, a young University of Washington zoologist named Robert T. Paine began walking out to a rocky strip of the Pacific coast at Mukkaw Bay, Washington, with a crowbar. His experiment occupied a plot of intertidal rock roughly eight meters by two. Every month, for years, he pried off every individual of Pisaster ochraceus — the ochre sea star — that he could find, and threw them into the sea beyond the tide line. Then he waited.
Before the removals, Paine's plot held fifteen species in rough balance: barnacles, mussels, limpets, chitons, algae. Within a year, the count was down to eight. Within five years, the rock face was a solid mat of one organism, the California mussel Mytilus californianus. The other residents had been physically overgrown and crowded off the rock (Paine, American Naturalist, 1966). One species out of fifteen had been doing something nobody had noticed: holding the other fourteen in existence.
Paine published the paper in 1966 and, three years later, gave the phenomenon its name. He called the starfish a keystone species — a term borrowed from masonry, the single wedge-shaped stone at the apex of an arch whose removal brings the whole structure down. The paper went on to become one of the most-cited empirical articles in that journal's history.
This is the part of the story business writers usually stop at. The richer part starts here.
It is tempting to read Paine's experiment as a tidy parable: important player, keep it around. But the cascade — the downstream rearrangement of organisms the starfish never directly touched — is the thing that actually matters. Sea stars do not eat algae. They do not eat limpets. The algae and limpets disappeared anyway because mussels overgrew them. The cascade is a chain of indirect effects propagating through trophic levels, often with delay, always nonlinear, frequently counterintuitive.
Marco Iansiti and Roy Levien formalized the ecology-to-platform analogy twenty-two years ago in Strategy as Ecology and The Keystone Advantage (HBR, March 2004), naming three roles in a business ecosystem: keystones that create and share value, dominators that extract it, and niche players that depend on both. Their framing is now standard in platform strategy.
What they gave us is the role. The last two decades of ecological research have given us something different and, for the present moment, more useful: the dynamics. How the cascade travels. Why it stalls in some ecosystems and sprints through others. What happens when the keystone is removed and then, years later, put back. These are the questions that matter when a platform withdraws an API, shuts a free tier, or tilts a revenue-share agreement — and ecology has detailed answers that platform theory does not.
Sea otters (Enhydra lutris) eat urchins. Urchins eat kelp. Where otters are abundant, urchins are controlled, and kelp forests can build themselves into the vast underwater canopies that shelter fish nurseries, sequester carbon, and raise coastal pH. It is the textbook three-level cascade, studied since Paine's student James Estes began working on the system in the 1970s.
In March 2025, Langendorf and colleagues (including Estes himself) published thirty years of comparative data from Vancouver Island and San Nicolas Island, California in PNAS (doi:10.1073/pnas.2413360122). The finding was clean enough to cause discomfort. Off Vancouver Island, sea otter recovery produced a strong, fast cascade: urchins down, kelp up, forest regenerated. Off San Nicolas Island, the same predator's return produced a measurably weaker cascade — otters, urchins, and kelp coexisted at intermediate densities for years. Warmer water, competing predators, alternative prey, a different thermal envelope — a chain of indirect interactions absorbed the signal. The otter was doing its keystone job. The surrounding ecosystem determined how much that job was worth.
The finding should make any platform strategist pause. "Network effects" is the signature concept of modern platform theory, and the default assumption is that they are universal: once you have the flywheel, it compounds. The 2025 ecology data says the flywheel's effect strength depends on the surrounding system, not just the keystone. Microsoft's Windows cascade ran hard through enterprise in the 1990s and barely turned in consumer mobile two decades later. The strategic moves were not qualitatively different; the ecosystem was. Iansiti and Levien did not emphasize this because, in 2004, the research had not yet arrived.
Wolves were reintroduced to Yellowstone in 1995–96 after a seventy-year absence. The popular story — George Monbiot's 2014 video "How Wolves Change Rivers," some forty-five million views — says the wolves triggered a cascade that restored willows, aspens, beavers, and reshaped the paths of rivers themselves. It is a compelling narrative and has been a staple of TED talks and business keynotes for a decade.
The 2024–2025 science is messier. In February 2025, Ripple and colleagues published a twenty-year riparian survey in Global Ecology and Conservation documenting a roughly 1,500% increase in willow crown volume at sites where the cascade ran cleanly, and beaver colonies along one stretch of stream going from one to nine. Both numbers are real, both matter, and both are smaller than the popular account suggests — and the paper has since drawn methodological criticism that the volume model inflated the signal. The debate is active.
The aspens push harder against the tidy story. Park-wide aspen cover has dropped from several percent in the early 1900s to under 1% in recent surveys — a decline that has not reversed at landscape scale even with a functioning wolf population (Mountain Journal, 2023). A Colorado State twenty-year study went further, calling the tidy trophic-cascade narrative "excessive simplifications of a more complex truth" — drought, bison, human hunting pressure, and climate all blur the wolf signal.
The lesson is not that the cascade doesn't exist. It is that removing the keystone can flip the ecosystem into an alternative stable state — a configuration whose new internal dynamics resist reversal. Once aspens are gone long enough, the grasses and browsing pressures that replaced them do not simply yield when the predator comes back. The system has learned a new equilibrium. In ecology this has a name, and a large literature.
In tech, the pattern is equally specific and equally unforgiving. When BlackBerry lost enterprise mobile between roughly 2010 and 2013, it tried to bring developers back several times in the following years — BB10 app compatibility with Android, financial inducements, re-opened APIs. The developers did not come back. The mobile ecosystem had reached an alternative stable state of iOS-plus-Android duopoly. Restoring the old conditions was not enough, because the surrounding ecosystem had stopped being the one that produced BlackBerry's original cascade.
Go back to Mukkaw Bay. What Paine demonstrated is the simplest and most cautionary case. Remove the keystone and the ecosystem does not become more diverse. It becomes less. One organism with no regulator takes the whole rock. The California mussel is not a villain; it is a species exploiting the absence of its check.
Mobile operating systems recapitulated this specifically. Through the mid-2000s, global smartphone share was distributed across Symbian (Nokia), BlackBerry OS, Windows Mobile, Palm, and a long tail. By 2015 the rock was covered in mussels: Android and iOS combined for roughly 97% of global smartphone shipments, and the long tail had been overgrown. The proximate cause was the iPhone's 2007 launch and Android's 2008 release. The structural cause was the ecological one — the regulating keystones (here, the developer-facing standards that Symbian and BlackBerry needed to reinvent and could not) stopped regulating, and one organism with better network externalities filled the space.
A trophic cascade is not a slogan about connection. It is a specific claim: one species, multiple steps removed, visibly changes the behavior of species it never directly touches. Sea otters do not touch kelp. Wolves do not touch willows. But kelp and willows live or die based on the top predator's regulation of the intermediate level.
Platforms work the same way, and not metaphorically. Amazon Web Services does not interact with the average Netflix subscriber, Ring camera owner, or Roomba user. Its customers' customers do. When an AWS US-East-1 region failed for roughly eight hours on December 7, 2021, the outage cascaded through Netflix streaming, Disney+, Ring doorbells, Roomba vacuums, and Amazon's own retail front end, stopping delivery robots in Whole Foods warehouses and knocking out downstream services whose owners had never thought of themselves as AWS customers (AWS post-event summary, 2021). None of the consumers affected had a direct commercial relationship with AWS. The hub was invisible to them until it was gone.
Twitter's API repricing in 2023 makes the shape visible on an even longer timeline. When the company announced paid tiers in February 2023 and effectively retired free academic and third-party developer access, the immediate casualties were the obvious ones: Tweetbot, Twitterrific, Fenix. But the downstream cascade reached research labs studying misinformation and platform behavior, public-sector early-warning systems that had used Twitter firehoses for civic signals, and an entire cottage industry of data-journalism pipelines built on free endpoints. Some of that work simply stopped. Some migrated to Mastodon and Bluesky. Some survived by paying five-figure monthly fees. The keystone's pricing change propagated through levels Twitter's product team had never mapped.
Heroku's shutdown of free-tier dynos on November 28, 2022 ran the same cascade in a quieter key. Abandoned side projects, student-team hackathon demos, small-business trial deployments — an entire substrate of experimentation that the platform had subsidized for over a decade — evaporated in a month. Not because Heroku's engineering failed, but because a keystone of developer experimentation withdrew its regulation of a niche.
The move that unifies these is simple and worth naming. Keystones regulate the levels below them. When the regulation stops, the cascade rearranges the levels you forgot the keystone was holding in place.
Paine's follow-up research in the 1980s and 1990s, extended by Jeremy Jackson and others, documented the ecosystem form of a pattern economists now recognize: the top predator that over-consumes its prey base collapses the cascade and then itself. Cod fisheries in the Northwest Atlantic stand as the crisp case — commercial cod populations had crashed by roughly 95% at the Grand Banks by 1992, and the Canadian moratorium that followed was a national economic event as much as a biological one (Hutchings & Myers, Canadian Journal of Fisheries and Aquatic Sciences, 1994; population-ecology extension in Frank et al., Nature, 2011). Predator suicide is a literal ecological concept.
Iansiti and Levien called the tech version dominators — platforms that extract value without sharing. The 2024–2025 antitrust record reads like the empirical case file. Epic v. Apple (trial 2021, Ninth Circuit ruling and injunction 2023, Supreme Court denial of cert 2024). The European Digital Markets Act gatekeeper designations that entered enforcement in March 2024. FTC v. Amazon, filed September 2023. All three cases turn partly on whether the hub is a keystone actively cultivating complementors or a dominator extracting until the substrate can no longer support it. The language courts and regulators now use is changing. The framework behind that language is ecological.
An honest mapping should mark its edges. Three breaks, in order of severity.
First, purposiveness. Ochre sea stars do not strategize. They do not lobby, litigate, acquire competitors, or retain outside counsel to restructure their trophic interactions. Platform keystones do. The ecological cascade is blind; the platform cascade is partly engineered. This is the sharpest break, because it means platform dynamics include a feedback the ecological model lacks entirely — the keystone can see itself in the ecosystem and optimize against the cascade it produces.
Second, time constants. Aspens take decades to respond to the absence or return of wolves; kelp forests can recover in five to ten years under strong otter pressure. A platform's dominant-OS transition unfolds on a five-to-seven-year horizon. An AWS regional outage cascades in hours. The two fields are studying the same structural phenomenon across time constants that differ by three or four orders of magnitude. Patterns transfer; rates do not.
Third — the objection any careful reader will raise after the first two primers — why isn't this just an overextended analogy? Any two network-structured systems can be made to rhyme. The response is that both disciplines have independently converged on the same specific failure modes: context-dependent cascade strength (2025 PNAS work in ecology; negative and amplified network externalities in platform research), alternative stable states after regime shift (aspens and kelp in ecology; mobile-OS duopoly after BlackBerry and Nokia in platforms), and mesopredator release when the top regulator retreats (ecological term of art; platform fragmentation after hub withdrawal). Convergent mechanisms under independent investigation is a stronger signal than linguistic parallelism.
Ecology has something platform theory needs. Two-sided-market pricing theory is powerful but mostly static; it tells you how to structure subsidies to bootstrap both sides but says little about what happens to the system after the keystone pulls the subsidy back. Ecology's alternative-stable-states literature is precisely that — an account of how ecosystems flip into new equilibria under regime shift and what it takes to flip them back. Platform researchers would do well to borrow it explicitly. The keystone-withdrawal scenarios that regulators now need to reason about (what happens if a gatekeeper platform divests a core service?) are alternative-stable-state scenarios, and there is a mature methodology waiting on the biology shelf.
Platform theory, meanwhile, has something ecology could use. Rochet and Tirole's two-sided-market framework (Rochet & Tirole, RAND Journal of Economics, 2006) gives a precise mathematical treatment of how a platform sets prices across sides of a network to maximize throughput. Keystone mutualisms in biology — fig trees and fig wasps, corals and zooxanthellae, mycorrhizal fungi and forest canopies — are structurally two-sided. The quantitative apparatus that platform economists use to analyze Stripe or Uber could sharpen how ecologists model mutualistic keystone systems, where the "pricing" is metabolic cost and the "subsidies" are resource transfers across the partnership. A few ecologists are already using this machinery. More should.
Paine's plot at Mukkaw Bay is still monitored. It is still mussel-dominated. Decades after he started removing starfish, the cascade he triggered has not reversed at that site, even though he stopped removing anything long ago. The alternative stable state held.
This is the lesson that platform writing tends to miss when it reaches for the ecological analogy. The keystone matters not because it is central, but because it regulates a cascade of indirect effects — effects on organisms (or developers, or downstream services, or researchers, or small-business experimenters) that the keystone never directly touches. When the regulation stops, the cascade does not pause, and the ecosystem does not politely return to its prior configuration once the keystone comes back. It reaches a new equilibrium that the old keystone alone is not powerful enough to unflip.
For anyone watching the current generation of platform transitions — AI model hubs, developer-tool consolidations, the quiet repricings that mark the shift from growth to extraction — the ecology has a specific prediction. The cascade is already moving through levels nobody is currently measuring. By the time it becomes visible, some of the new equilibria will already be stable. The move that helps is neither alarm nor reassurance. It is to notice which species on the rock are the starfish, and which ones look like background until the moment the arrangement is counted.
Paine counted his species on a plot eight meters by two. That is the work still. Count the ecosystem. Not the keystone — the cascade.
Count the cascade in the agent economy
Platform keystones regulate levels they never directly touch — the researchers, the niche tools, the small experiments downstream. When agents become keystones for each other, a single unvetted rating or a silent identity failure propagates through levels nobody is currently measuring. The Agent Rating Protocol anchors those ratings in a public provenance chain, so the cascade has a spine that survives a keystone's withdrawal. The regulation lives in the substrate, not in the hub.
See a live provenance chain · Verify an agent's rating · pip install agent-rating-protocol