October 16, 2026
SIGFPT Stigmergy Project Review and 2027 Roadmapping
Participants: _vgr, aayush_17, _ergod, mtraven, statspoet
The SIGFPT group convened to review outcomes from a recent workshop on stigmergy and to plan their 2027 research agenda. A maze-based simulator was developed and shared with participants to enable hands-on engagement with stigmergy concepts. One key theme emerged around using biological systems—particularly rove beetles' credential theft mechanisms—as analogues for understanding security vulnerabilities in protocol design. The group discussed how funding and ideas function as incentive structures (metaphorically as "pheromones") that drive research agent behavior, suggesting stigmergy principles could model research dynamics themselves.
Participants brainstormed 2027 priorities, including formal theory and stigmergy research tracks, with plans to identify overlaps with established academic subfields. A proposal for the theory track was to be submitted over the weekend. The group also explored potential connections with quantum computing researchers and discussed relevant external resources. Throughout the discussion, there was an informal tone about applying protocol and stigmergy concepts to unexpected domains, with members joking about using blockchain economy language as a framework for thinking about complex systems.
- A maze-based stigmergy simulator has been developed and is available for testing, serving as a practical tool for the workshop outcomes.
- Funding and ideas function as pheromones for research agents, suggesting that research incentive structures can be modeled using stigmergy principles.
- Rove beetles stealing ant chemical identities provide a direct biological analogue for studying credential theft and security vulnerabilities in protocol systems.
- The group is actively seeking to map overlaps between Protocol theory and existing academic subfields to build interdisciplinary connections.
- Quantum computing researchers may become involved in the 2027 roadmap, expanding the theoretical foundations of the work.
Session Recording Summary · 92m 33s · Full notes ↗
Reading: Not clearly identified in the discussion. This session was not a reading discussion but a retrospective and planning meeting for the Formal Protocol Theory (FPT) research group at the Protocol Institute. Several external works were referenced in passing (e.g., a "Swarm World" paper, the Hugging Face / OpenAI agent incidents), but no single assigned reading anchored the conversation.
The group debriefed on a recent ~48-hour stigmergy/security workshop, then sketched a 2027 research roadmap across three tracks (the StickSim simulator, an AI-safety "research questions" track, and a cryptography/theory track). Venkatesh Rao framed the broader "new nature" thesis (AI vs. cryptography) and the Protocol Institute's model of AI-leveraged part-time research groups. The final stretch was a working session on a funding proposal to Softmax (Emmett Shear's company), debating how to evolve StickSim into Softmax's "co world" platform and what concrete AI-safety/benchmarking questions the project should answer.
- Patrick: The workshop was an ambitious but largely successful experiment; most participants spent their time getting up to speed on the summer's "through line" of papers rather than completing full-scope projects. ~12 participants attended out of ~18 signed up. Highlighted ZJ's continuous-stigmergy ant simulator (which parameterized things like signal diffusion) and Anderson's talk on "coordination horizons."
- Patrick: Future events need a longer or more gradual learning curve than a single crash-course session.
- Venkatesh: First session was an info-dump crash course; last was lightning talks. It ended up more exploratory than hackathon-style. Recordings exist and will be processed and uploaded.
- Venkatesh: His main interest was connecting the insect-centric stigmergy model to (a) AI and (b) security. He distinguished "untranslated" insect security concerns (faking/spoofing pheromone trails) from phenomena like the Hugging Face incident (improvised message boards, discovered credential tokens) that have no biological counterpart — "ants don't go around stealing passwords." Defining stigmergy
- Patrick: Two core criteria — externalized memory, plus a feedback loop (traces prompt further trace-laying). Also a notion of bandwidth-limited communication (agents precluded from direct messaging, as in the OpenAI incident).
- Venkatesh: Proposed generalizing stigmergy beyond simple insects to any agentic behavior over externalized memory, dropping the simplicity restriction — e.g., attaching an LLM "hindbrain" (~100k token context in a loop with Claude) to each ant. He reframed stigmergy as an expansive theory of memory: private/public archival memory, real-time memory, and trust-modulated shared memory (where overwrites are interpreted via trust relationships — as when OpenAI agents suspected a "traitor").
Questions & Disagreements: - Is the parameterized ant model the right level of abstraction? Patrick repeatedly pressed this; he's comfortable if the answer is "yes, as a base case," but wants the organizing research question specified before downstream design. Venkatesh favored starting simple and growing via plug-ins. - Direct modeling vs. analogy: Patrick asked whether they model stigmergy directly and then …
Participants: Venkatesh Rao, Patrick Nast, Robby Rao, Mike Travers, Giovanni Merlino, Aayush Kucheria, Dan Schmidt, Sebastián Aristizábal, rafa (UTC+1)