AI Nature Company · Journey series
Papers and references
Papers the AI Nature Company founding chat treated as useful, with titles verified from the papers and one line on why each one matters here.
These are the references the 2026-08-29 founding conversation treated as useful. Titles were checked against the papers themselves. ChatGPT tracking URLs were stripped.
One line each is enough. The papers are not a reading list for its own sake. They are the reasons the design has gates.
Scaling and verification
Towards a Science of Scaling Agent Systems
Kim, Gu, Park, et al. Google Research / MIT / Google DeepMind.
Why it matters here: architectures without centralized verification propagate errors; coordination has a cost. The MCP substrate is the verification layer. Adding more agents is not the plan.
Failure modes and topology
Why Do Multi-Agent LLM Systems Fail?
Cemri, Pan, Yang, et al. UC Berkeley. MAST taxonomy.
Why it matters here: specification issues, inter-agent misalignment, and weak task verification show up as soon as you have more than one agent. Roles exist to keep those modes from becoming the company.
Evolution through a commit gate
Position: Agentic Evolution is the Path to Evolving LLMs
Lin, Lu, Shi, et al.
Why it matters here: evidence should become a candidate update, then pass an explicit commit. The nested loops put that gate on daily memory and keep reflex from rewriting the constitution.
Memory can make the work worse
On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification
Ye et al. Salesforce AI Research, August 2026.
Why it matters here: memory-based self-improvement shows high run-to-run variance and depends on task order. A loop that looks like learning can be degradation. Immune review and the founder kill switch are the answer, not a bigger memory bank.
Colony coordination, as analogy
Noise improves collective decision-making by ants in dynamic environments
Dussutour, Beekman, Nicolis, and Meyer. Proceedings of the Royal Society B, 2009.
Why it matters here: a colony can coordinate through traces in a shared environment. That is the stigmergy picture behind a market of cells. It is an analogy. It does not replace identity, audit, or a human founder.
How to use this page
When a later dispatch changes the design, it should say which paper stopped applying, or which new paper earned a line. Do not add a citation for atmosphere. If it did not change a gate, a role, or a stop rule, it does not belong here.