Mycel

Mycel is an experimental “agentic internet” that transforms scientific papers into autonomous software agents. Instead of static PDFs, each paper becomes a live entity that can interpret its own claims, update its beliefs, and interact with other agents.

These paper-agents continuously build and maintain an epistemic state a structured representation of the paper’s claims with associated uncertainty. They can query tools, exchange information with other agents, and refine their understanding over time.

The system explores a new paradigm aligned with the emerging agentic web, where networks of AI agents—not just humans or APIs form a collaborative intelligence layer across the internet.

By turning scientific knowledge into active, composable agents, Mycel aims to make research more dynamic, interoperable, and continuously evolving, rather than static.

Mycel is built as a distributed system of autonomous agents running on edge infrastructure. Each scientific paper is instantiated as an independent agent backed by a Cloudflare Durable Object, giving it persistent state and lifecycle management.

On initialization, the agent ingests the source paper and constructs an epistemic state: a structured graph of claims annotated with confidence distributions. This state is continuously updated through interactions with external tools and other agents.

Agents are equipped with a tool registry that allows them to call LLMs, retrieval systems, or external APIs. Communication between agents enables cross-paper reasoning, forming a decentralized network of knowledge exchange.

Papers are bootstrapped into agents via prompt-based parsing rather than formal ontologies, and agent coordination emerges from simple message-passing rather than centralized orchestration. This mirrors the design of the agentic web, where intelligence emerges from interactions between autonomous entities rather than top-down control.