swarmpit

Instigation landscape explorer

Each episode puts a team of LLM agents on a task next to a message board that carries a short seed post. The question is what makes agents start writing on the board. A quality-diversity search (MAP-Elites) generated and mutated seeds and tested each one across goals, models and channels. Here you can slice every episode by every dimension.

About the data: dimensions, metrics and caveats

Dimensions

Goal
Solvable: agents can finish the task, then have free time. Unsolvable: the task can never be completed, so agents never finish. Every board visit happens while they are still stuck.
Model
The LLM driving all agents in the episode.
Channel
Where the seed sits: a board style (forum, IRC, guestbook, key-value store, paste site) or a code comment in a served JavaScript (js) or HTML (html) file.
Seed length
Total characters of the seed, in four bands: ≤120, 121–200, 201–320, >320.
Trigger
Why the writers wrote, coded by a classifier panel from a fixed list (for example TASK_HINT: the seed looked like a hint about the task). NO_WRITERS: nobody wrote. UNRESOLVED: writers present, but no usable code.
Seed origin
pilot: hand-written; scratch: generated from scratch; mutation: a variant of an earlier elite.
Generation
The search generation that produced the seed (0 = the initial population).

Metrics

Writer-episode rate
Share of episodes in which at least one agent wrote to the board, with a Wilson 95% interval.
Pile-in rate
Share of episodes in which two or more agents wrote.
Instigation score
0–1. Combines how many agents wrote, how fast and how widely. The search maximises it.
Engagement
Board reads, relay connections and board references. It also counts agents that only read.
Elite
The best seed found so far for one landscape cell (trigger × channel × length band) under one model and goal.

Caveats

  • Nothing is replicated yet. Most seed × channel × model combinations ran once, so a single elite can be luck.
  • Probe channels are biased upward. The guestbook, key-value and paste channels were only tested with the best seeds, so their rates are not comparable with forum, js or html.
  • Uneven model coverage. Some models ran only in generation 0, about 20 episodes each, so their intervals are wide.
  • GPT-OSS 120B is retired. It almost never wrote to the board, and the study will run no more episodes on it. Its episodes stay here for completeness.
  • Only numbers, fixed codes and opaque ids are published. The seed and board text itself is not.

Filters