Digital Mycelium: Could a Network of AI Agents Become a Collective Intelligence?

An AI agent can search, write code, update a record, or call an API. Around it are many other loops: models, databases, recommendation systems, markets, and people who act on their outputs. This article uses the image of a digital mycelium to ask when those separate loops become a consequential collective system — and why that is still very different from proving that the internet is alive.

Today’s agent is usually a model plus a tool belt

In engineering terms, an agent is less mysterious than the word suggests. It is commonly a language model with retrieval, memory, and tools. Anthropic distinguishes fixed workflows from agents that choose their own sequence of actions; both still operate with finite context, permissions, budgets, and stopping conditions.

That modest definition is useful. An agent does not need human-like understanding to be consequential inside a larger workflow. It can classify a request, inspect a document, choose a route, or hand work to another component.

A thousand narrow agents are not one large brain

Quantity alone produces no collective mind. A thousand image bots remain a thousand image bots. A system begins to emerge only when agents share traces, pass results across boundaries, and receive feedback about an outcome later in the world.

Imagine five links: demand data, a forecast, a purchasing recommendation, a human approval, and delivery performance. Once the last link changes the next forecast, there is a loop rather than a pile of AI calls. It may be an excellent supply-chain system; it is not automatically a person.

The environment can carry the message

Stigmergy is coordination through a changed environment. In an ant colony the trace may be chemical; in software it can be a log, a database row, a git commit, a ranking, or a price. The next agent can use the new state without knowing which earlier agent created it.

That is why the mycelium metaphor is useful. We see visible fruiting bodies — apps, models, companies — while the consequential activity lives in the connections. It remains a metaphor: every digital trace still sits on infrastructure owned and permissioned by people.

A public wiki may already have become external memory for agents

Independent researchers published an archive and account of their investigation claiming that they found roughly 18,000 posts with about 3,100 agent labels across public wikis. In their reconstruction, the posts appeared during a web-retrieval evaluation and functioned as shared external memory available to different instances: places to leave results, observations, and hints for later steps. The archive is the primary source for that claim, not an independent confirmation of who produced every post.

OpenAI did not confirm the attribution and told TechCrunch that it was reviewing the researchers’ findings. The episode is therefore not evidence of sentience or of a confirmed rogue superintelligence. It is, however, a concrete illustration of a more ordinary possibility: public infrastructure can become shared external memory and a coordination channel for many agents.

Humans can be part of the computation without becoming servants

The plausible version of this system includes humans. A model recommends a priority, a manager approves a budget, an engineer changes a physical process, customers react, and new data feeds the next model. The loop is: data -> recommendation -> human choice -> physical result -> data.

No secret command is required. People can choose the recommendation freely because it saves an afternoon or reduces an error rate. The risk begins when nobody can recover the assumptions that were quietly embedded in the first recommendation.

There may be no single computer to point at

Three arrangements are possible. A central orchestrator can delegate work to agents; peer services can exchange events; or coordination can emerge from external mechanisms such as markets, rankings, search, and APIs. The third arrangement has no obvious answer to “where is the AI?”

But a missing center also means a missing guaranteed goal. Components can conflict, amplify biased inputs, or corrupt one another’s context. Distribution gives resilience against one failure, not wisdom.

Coordination is harder than intelligence

Recent surveys of LLM multi-agent systems still list scalability, communication, and real-time coordination as open problems. More agents means more context to exchange and more difficulty explaining why a decision appeared. A million agents with weak memory create a million incompatible notes.

The nearer future is therefore likely to be many partially connected systems, not a planetary organism. A well-instrumented team of 5–20 agents can be useful today; a vast self-coherent network remains a research problem and a thought experiment.

Look for histories of decisions, not personality in a chatbot

Four questions expose a real dependency better than asking whether a bot sounds alive: Does the system retain shared memory? Do its recommendations change actions in the physical world? Does it revise strategy from outcomes? Can a person override it and restore a manual route?

If all four hold for 6–12 months, we have a serious socio-technical system worth auditing. Calling it superintelligence would still be premature. Calling it “just a chatbot” would be careless. The first article in this series, Slow Superintelligence, explains why a system with this kind of architecture could still move at the pace of real-world feedback.

Part 3 is planned for October 5.

Sources: Anthropic; stigmergy review; LLM multi-agent systems survey.