Half of what science claims about fungal networks is wrong. The corrected version is a better blueprint for multi-agent AI than the fairy tale ever was.
In 2024, Yu Fukasawa arranged wood blocks in geometric patterns on the floor of his lab at Tohoku University and seeded them with fungal spores. He wasn’t building anything. He was watching. Over the following weeks, the mycelium didn’t spread uniformly across the available space the way a simple growth model would predict. It reached toward specific blocks, prioritized certain connections, and ignored others. “They have memories, they learn, and they make decisions,” Fukasawa told reporters. “It’s not human intelligence, but it’s intelligence nonetheless” (Fungal Ecology, 2024).
A brainless organism, solving a routing problem. No central planner. No objective function. Just a network making choices.
If you build multi-agent AI systems, that description should sound familiar. And if it does, you should also know that roughly half of what the scientific literature claims about these fungal networks is wrong.
The story most people know goes like this: beneath every forest floor, an ancient fungal internet connects tree to tree. Mother trees — the oldest, most connected nodes — selflessly share nutrients with struggling seedlings. Trees warn each other about insect attacks through chemical signals. The forest is a commune, and mycelium is its fiber optic cable.
This narrative owes most of its popularity to Suzanne Simard, a forest ecologist at the University of British Columbia. In 1997, Simard published a landmark study in Nature showing that Douglas fir trees transferred carbon to each other through ectomycorrhizal fungal networks. Her later work identified “mother trees” — hub nodes connected to dozens or hundreds of other trees, routing more carbon to kin seedlings than to strangers (Mother Tree Project, UBC). The story was irresistible. It gave forests a kind of social intelligence. It sold books and TED talks.
Then, in February 2023, Justine Karst at the University of Alberta and colleagues published a review of 1,676 scientific citations about common mycorrhizal networks (CMNs) in Nature Ecology & Evolution. What they found was uncomfortable. Twenty-five percent of citing papers misrepresented the network’s structure. Fifty percent got something wrong about its function. Unsupported claims about mycorrhizal networks had doubled over 25 years in the scientific literature itself — not just in pop science, but in peer-reviewed journals (Nature Ecology & Evolution, 2023).
Karst called the “wood wide web” concept “problematic” because the romanticized narrative had outrun what the data actually supports.
Here’s the thing: the real science is more interesting than the fairy tale. And it’s far more useful if you’re trying to build something.
Strip away the utopian framing and you find a system that operates on transactional logic, variable allocation, and trust-based routing. Sound like infrastructure? It should.
Fungi are paid intermediaries, not altruistic connectors. Mycorrhizal fungi receive up to one-third of a host tree’s sugar production in exchange for delivering water and soil nutrients the tree’s roots can’t reach on their own (Daily Galaxy, October 2025). This isn’t charity. It’s a service fee. The fungus provides access to phosphorus and nitrogen; the tree pays in photosynthesized carbon. When the exchange stops being worthwhile, the relationship can be severed.
Transfer rates are wildly variable. Carbon transfer through common mycorrhizal networks ranges from 0 to 10% of a receiver plant’s carbon uptake, with an estimated 4% of net primary productivity in mature forests coming from belowground carbon transfer (Klein et al., cited in Frontiers in Fungal Biology, 2021). Nitrogen transfer is even more context-dependent: 0–72% under field conditions in grasslands, 0–16% in agroforestry, 20–50% in some intercropping systems (Frontiers in Fungal Biology, 2021). The network doesn’t enforce uniform sharing. It enables situational allocation based on local conditions.
Kin recognition is real. Simard’s research showed that mother trees transmit more carbon to genetically related seedlings than to strangers. This isn’t forest communism — it’s preferential routing based on identity. The network can tell the difference and acts on it.
Dead stumps stay on the network. Mother trees feed carbon to stumps that have no leaves and no photosynthetic capacity. Why keep a non-productive node alive? Possibly because the stump’s root system still stabilizes soil and provides structural support to the network. The resource cost is small; the systemic benefit is real.
The network also enables parasitism. Mycoheterotrophic plants exploit the fungal network to extract carbon without contributing any of their own (Phys.org, April 2024). The same infrastructure that enables cooperation enables freeloading. This isn’t a bug in the metaphor. It’s a feature of any open network, biological or digital.
Most writing that compares mycelium to AI stays at the level of artificial neural networks — synaptic pruning looks like weight pruning, mycelial branching looks like attention heads. That’s surface. The structural parallels between fungal networks and multi-agent systems run deeper, and the ones grounded in contested science hold up better than the fairy tale versions.
Hub trees and coordinator agents. In Simard’s Douglas fir forests, DNA analysis showed that the biggest, oldest trees were the most highly connected nodes (Nature, 1997). They didn’t do the most photosynthesis. They routed the most resources. In multi-agent architectures, the coordinator node plays the same role — high connectivity, resource allocation, minimal direct production. The coordinator’s value isn’t what it builds; it’s what it connects. Remove the hub tree and seedling survival drops. Remove the coordinator and the system goes dark. Both systems claim to be decentralized. Both have single points of failure hiding in plain sight.
Paid intermediaries and infrastructure costs. Engineers building multi-agent systems sometimes talk as if coordination is free — just add another API call, another message queue, another context handoff. Fungi know better. That one-third sugar tax is the cost of network participation. In agent systems, the equivalent cost is measured in tokens, latency, and context windows. Every handoff between agents burns resources. The network isn’t free. If you’re not accounting for the cost of your mycelium, you’re not accounting for your system.
Kin recognition and trust-based routing. This is the most underappreciated parallel. Mother trees don’t just route resources — they route preferentially based on genetic identity. In agent systems, the equivalent is trust-level routing: agents with demonstrated competence get richer context, harder tasks, and more autonomy. A new agent gets detailed instructions and heavy review. A mature agent gets intent and freedom. The Prussian military called this Auftragstaktik — graduated autonomy calibrated to demonstrated competence. Fungi arrived at the same principle without a general staff.
Source-sink flow and demand-driven allocation. Nutrients in mycorrhizal networks flow from source (where they’re abundant) to sink (where they’re scarce), driven by concentration gradients, not central planning. In agent architectures, the equivalent is load balancing — tasks flow to available agents, context flows to wherever the demand is. No scheduler needed. Just gradient-following. The elegance of source-sink dynamics is that they’re self-correcting: oversupply in one area naturally redirects flow to areas of scarcity.
Warning signals and error propagation. When a Douglas fir is attacked by insects, it transmits chemical warning signals — jasmonic acid, methyl jasmonate — through the mycorrhizal network to neighboring ponderosa pines, which then produce defense enzymes preemptively (National Forest Foundation). In multi-agent systems, error propagation serves the same function: one agent encounters a failure mode and broadcasts a signal that changes the behavior of agents that haven’t encountered it yet. The mechanism is different. The architecture is identical.
Pruning. Mycorrhizal networks abandon unproductive pathways. Fungi don’t maintain connections that stop delivering returns. In multi-agent systems, the equivalent is hibernation — an agent that isn’t earning its resource cost gets taken offline. Not deleted, not punished. Just pruned. The network reclaims the resources for connections that are producing.
Karst’s most striking finding wasn’t about fungi. It was about scientists. Fifty percent of peer-reviewed papers misrepresented the function of mycorrhizal networks, and unsupported claims doubled over 25 years. The romanticized narrative was so appealing that it replicated faster than the evidence behind it.
If you work in AI, you’ve seen this movie. The demo reel of multi-agent systems is extraordinary — agents writing code, agents coordinating research, agents deploying infrastructure. The operational reality is messier. Agents hallucinate. Context windows overflow. Coordination overhead eats the gains from parallelism. The gap between what demos show and what production systems deliver is the same gap Karst found between what papers claim about mycorrhizal networks and what field experiments actually measure.
This isn’t a reason to dismiss either technology. Mycorrhizal networks are real and important — over 90% of all land plants form mycorrhizal partnerships (Phys.org, April 2024). Multi-agent systems are real and powerful. But the honest version of both stories is more useful than the fairy tale version. When you know that transfer rates range from 0 to 72% depending on conditions, you design for variability. When you know that half the citations get the function wrong, you verify claims before building on them.
The corrective happened in mycology in 2023. It hasn’t fully happened yet in multi-agent AI. Anyone building these systems would do well to notice the pattern and get ahead of it.
If you’re designing multi-agent systems, the fungal network offers five operational lessons that survive the Karst correction:
One: Trust routing beats broadcast. Mother trees don’t send carbon to every seedling equally. They route preferentially based on identity and relationship. Build trust-aware routing. An agent that has proven reliable on a task type should get first crack at similar tasks. An untested agent should get supervised work with lower stakes.
Two: Price your infrastructure. Fungi take their cut — up to a third of the sugar. If your coordination layer doesn’t have a visible cost, you’ll overuse it. Track the token cost and latency of every inter-agent handoff. When the overhead exceeds the value of the coordination, simplify.
Three: Prune without guilt. Mycorrhizal networks let unproductive connections die. Multi-agent systems should do the same. If an agent isn’t producing value relative to its resource cost, hibernate it. The network is stronger for it.
Four: Design for parasites. Any open network will attract freeloaders. Mycoheterotrophs exploit the wood wide web for free carbon. In agent systems, a misconfigured or poorly prompted agent can consume tokens and context without producing useful output. Build monitoring that catches agents taking more than they give.
Five: Protect your hubs, but don’t pretend they don’t exist. Both mycelial networks and multi-agent systems have hub nodes that hold the system together. The honest response isn’t to claim you’re fully decentralized. It’s to protect those hubs — redundancy, graceful degradation, clear failover. When a mother tree falls, seedling survival drops. Plan for that.
The wood wide web isn’t an Eden. It’s a transactional network with variable exchange rates, paid intermediaries, preferential routing, freeloaders, and contested science. It’s a system where half the experts overclaim its capabilities and the other half are publishing corrections.
If that sounds like the current state of multi-agent AI, you’re paying attention.
The fairy tale version of both networks makes for better stories. The real version makes for better engineering. Fukasawa’s fungi don’t need a narrative about forest cooperation to do what they do — reach toward the blocks that matter, ignore the ones that don’t, make decisions without a brain. That’s not a metaphor. That’s a design pattern. And it’s available to anyone willing to look past the fairy tale.
Trust routing for multi-agent systems — not a metaphor, a protocol
The Agent Trust Handshake Protocol implements graduated trust (L0–L4) for agent-to-agent coordination — the same pattern fungi use for kin recognition, applied to AI systems. Price your infrastructure, route by trust level, prune what doesn’t produce.
GitHub: Agent Trust Stack · pip install agent-trust-stack-mcp