Mycelitree maps millions of agents

Resonnence Labs’ Mycelitree is being discussed this week because it puts a concrete number on a problem many AI infrastructure teams have described more vaguely: too many agents, tool calls and traces to inspect as lists. A May 2026 post by the account mergedelta said Mycelitree can visualize about 1 million live agents on consumer hardware and could scale to 72 billion agents on upcoming NVIDIA rack systems. The company’s public site describes Mycelitree as an MCP server that “measures the space” between a command-line interface and its tools, with client-side computation and features aimed at drift, confidence and review independence. ### Why are people calling observability a spatial problem? The mergedelta post framed the issue as geometry rather than logging alone: once agent fleets become large enough, operators need ways to see structure, clustering and movement across systems, not just read one trace after another. That framing lines up with the language on Mycelitree’s site, which says the product is designed to surface drift, flag “performed confidence,” and show when multiple reviews are less independent than they appear. (mycelitree.com) Mycelitree’s public materials do not, in the fetched pages, spell out the rendering method behind the visualization claim. What they do show is a product pitched around relationships between outputs, tools and reviewers, rather than around a conventional dashboard of per-request logs. That makes the “spatial” description an inference from the product’s own wording and the social post, rather than a published technical paper. (mycelitree.com) ### What exactly is the 1 million versus 72 billion claim? The social post cited two thresholds: about 1 million live agents on consumer hardware now, and 72 billion agents on upcoming NVIDIA racks. The first number is presented as a current demonstration claim; the second is presented as a projection tied to future hardware rather than a deployed production count. NVIDIA’s published specifications help explain where the second number comes from, even though NVIDIA does not mention Mycelitree. (mycelitree.com) NVIDIA says its GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design, and that the 72-GPU NVLink domain acts as a single massive GPU. NVIDIA’s DGX GB200 materials describe the same 72-GPU rack architecture for trillion-parameter model training and inference. ### Does Mycelitree itself verify those numbers publicly? Mycelitree’s public homepage, as fetched, does not independently document the 1 million-agent demo or the 72 billion-agent projection in technical detail. The homepage instead markets the product as an MCP server, says all computation runs client-side, and lists Resonnence Labs as the company behind it. That means the scale figures should be read carefully. (nvidia.com) The 1 million figure is currently sourced to the social post highlighted in the briefing, while the 72 billion figure appears to be a hardware-linked projection rather than a benchmark published in vendor documentation. NVIDIA’s materials verify the rack configuration, not Mycelitree’s agent-count math. ### Why does this matter for production agent systems? (mycelitree.com) Google’s announcements this week that users will be able to create and manage multiple AI agents from Search add context for why observability vendors are emphasizing scale. Google said Search is adding agent features, and CNET’s I/O recap said users will be able to create and manage multiple AI agents directly from the search box. More agents in mainstream interfaces means more runtime actions, more tool calls and more states to watch. (nvidia.com) For infrastructure teams, the Mycelitree pitch is that monitoring large agent fleets may require new aggregation and visualization primitives, especially if one microservice call can correspond to one visible node or actor in a larger system. The next verifiable milestones are likely to be technical docs, demos or benchmarks from Resonnence Labs that show how the system renders, aggregates and updates those agent maps at the scales now being claimed. (nvidia.com) (mycelitree.com)