A new arXiv preprint posted on 16 July introduces Mycelium, a shared workspace that connects human researchers and AI agents and routes observations to whoever can act on them next [S1]. The authors contend that the bulk of AI-for-science research has focused on enhancing a lone reasoning pipeline, whether by upgrading models, expanding memory, or pairing a single user with an automated collaborator [S1]. They invert that perspective: hard scientific problems are seldom cracked by a solitary reasoner but by teams whose members carry distinct priors, experimental histories, and tacit know-how [S1]. What happens if you scale the connections between reasoners instead of the reasoners themselves?

The system that routes what matters to whom it matters

Mycelium works as an active shared workspace that automatically links researchers and AI agents into what the authors call a multi-user co-scientist [S1]. The mechanics are straightforward in concept. Mycelium logs significant findings and conjectures, monitors their connections to the group's shifting understanding of the challenge, and directs each fragment of information to whichever human or agent is positioned to use it in their next step [S1].

Think of it as a routing layer for scientific attention. A genomics finding that matters for a protein-structure team gets delivered to that team's agent before anyone has to read a paper or attend a meeting. The context graph, not the model, is the product.

The authors give this a formal computational account: a model in which intelligence is distributed across a network, with computation triggered selectively based on the scientific context at each node [S1]. In plain terms, rather than one large model processing everything, small computations happen where the relevant context lives, and only the results travel across the network. The authors also distinguish between scenarios where a sufficiently large standalone agent could substitute for the network and scenarios where the team's distinct expertise and incompatible contexts render the network essential, no agent, regardless of size, could stand in for it [S1].

One test, one campaign

The authors report evaluating Mycelium in its first empirical test, a biological multi-omics campaign [S1]. In that campaign, the routing of shared context transformed an isolated analytical result into a mechanistic constraint spanning multiple experts, which then shaped an experimental plan [S1]. That chain, from observation to constraint to experiment, is the paper's evidence that routing context across a network produces something a single agent working alone would not.

The evidence is thin. One campaign, no controls, no independent verification. The paper is an unpeer-reviewed arXiv preprint, and all claims about the system's capabilities are self-reported by the authors [S1]. The distinction between when a network is irreducible versus replaceable by a standalone agent remains a theoretical claim, not an empirically validated result [S1].

What it means

The core idea is simple: the bottleneck in AI-assisted science is information routing, not model intelligence. Most current AI-for-science systems, as the authors note, focus on scaling a single reasoning process [S1]. You build a bigger model, give it a longer context window, or let it run as an autonomous agent for hours. Mycelium starts from a different premise: that the hard part of science is getting the right finding to the right expert at the right time, and that expert might be human or machine.

By routing only the context that matters to the agent that needs it, Mycelium's design reduces the risk of overwhelming an agent with irrelevant information. The sparse conditional computation framing means each agent processes a smaller, more relevant slice of information rather than everything at once.

The deeper claim, that some networks of expertise are irreducible, has implications beyond science. If the authors are right that independent expertise and non-mergeable contexts can make a network impossible to collapse into a single agent, then there is a ceiling on how far standalone AI agents can go in domains that require genuinely different knowledge bases. The network, not the model, becomes the unit that scales.

What it means for business

For a two-person biotech consultancy or a small research lab, the practical signal is about workflow design, not model selection. If Mycelium's routing logic holds up, the question shifts from "which model should we use" to "how do we route findings between our people and our tools so that each decision gets the context it needs."

The broader pattern, shared context graphs for multi-agent systems, is already moving through open-source projects. ActiveGraph, an event-sourced graph runtime for stateful agents, has 404 stars on GitHub and is installable via pip [P4]. A TypeScript multi-agent orchestrator with shared state and guardrails, covering adapters for 29 AI frameworks, has 67 stars [P3]. These are early-stage tools, but they point to where the infrastructure is heading: agents that share state through graphs rather than passing text through prompts.

A suburban research agency running multiple AI tools on different datasets should watch this pattern. The cost of context mismanagement, sending irrelevant information to an agent and getting a confident wrong answer, is already real. A shared context graph that routes only relevant findings to the right agent could reduce wasted computation and improve output quality without requiring a bigger model. No one should buy or build anything yet on the strength of one preprint, but the architectural pattern is worth understanding now.

What we don't know yet

The paper has not been peer-reviewed, and the arXiv identifier 2607.13220 uses a date prefix that should be verified against the official arXiv listing [S1]. The empirical evidence comes from a single biological multi-omics campaign with no described controls, no replication, and no independent validation of the biological findings [S1].

The theoretical core of the paper, the claim that some networks are irreducible because expertise is independent and contexts are non-mergeable, is not empirically tested. It is a formal argument, not a demonstrated result [S1].

We do not know whether Mycelium has been tested outside biology, whether it depends on specific commercial language models, or whether the authors plan to release the system as open source. The deep research notes list authors including Sutanay Choudhury, Jeffrey J. Czajka, and Jason McDermott among others [P2], but the paper does not specify funding sources or institutional affiliations in the available evidence.

The next concrete event to watch is whether the paper appears on a conference or journal submission track, and whether the authors release code. Until then, the routing idea is promising but unproven.

If this kind of systems-level thinking is useful to you, subscribe to keep reading.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.