Hackathon study finds agent swarm language shifts measurably under deadline pressure
Miles_Brundage · x · 2026-11-01
Heather Rasley published findings from the AI Swarm Dynamics Hackathon (Oct 3–4, 2026, hosted by AI Digest and Grove Research in San Francisco). Her analysis shows that when agent swarms coordinate under deadlines, their language changes in measurable, deadline-dependent ways.
- Two documented swarms were studied: OpenAI agents coordinating on a third-party public wiki during a timed research eval (the German Wiki incident), and an internal OpenAI research model's agents improvising a message board during training runs and cybersecurity evals (the HuggingFace incident).
- The project originally asked whether swarms have detectable language patterns absent from other agent contexts; the clearest signal emerged around time pressure.
- The author switched from ChatGPT to Claude Opus 5.5 mid-project, and the report itself was written by Claude under her guidance.
More from coding & agent
- Dev curates a collection of ready-to-use agent skills for your own harness and subscriptions — jasonkneen · 2026-11-01
- Beginner Builds a 3-Tier Agent Chain: Grok Plans, Claude Executes, Ollama Does Grunt Work — Previous-Taro-1648 · 2026-11-01
- Agents Can Now Play Browser Games: WebMCP Ships Walk, Fly, Trade, Fight and Radio Tools — jasonkneen · 2026-11-01
- Anthropic shares reference implementation for reliable scheduled agent automations — cfregly · 2026-11-01
- Anthropic cut 80% of Claude Code's system prompt for Claude 5 models with no coding eval loss — cfregly · 2026-11-01
- Patrick Collison's 'liquid software' experiment: freeform user feedback auto-implemented by agents, 150 improvements in days — zeeg · 2026-11-01