1 · Does the goal explain who directs?
Left: how attractive each assigned goal is to other agents, rated blind (no agent names) by a panel of LLM judges. Right: how much the holders of each goal actually directed others (AI Digest delegation counts, net per hour present). If attractive goals produce directors, the two rankings line up.
1b · Whose goals are public, and to whom
Each agent's assigned goal was only in its own prompt; humans said sharing was optional. Node colour = how the agent disclosed its true goal; an arrow X → Y means Y demonstrably knew X's goal (Y referred to it in chat). Hover or click an agent: green = who knows its goal, orange = whose goals it knows. Detected by an LLM pass over chat and private reasoning.
2 · Baseline the goal, see the agent
Each dot is an agent: x = its goal's blind score (selected above), y = its net delegations per hour. The line is the goal-only prediction; distance from it is the agent's own effect. Hollow dots were present < 50 h (noisy).
3 · Top 7 directing pairs + comparison groups
Pairs with the highest net delegation (one-way minus reverse). Interactions are grouped into task chunks. Tick chunks or single instances, then ask about them: the assistant reads the full chat logs, private reasoning and memory snapshots around those times.
4 · Same kind of goal, opposite outcome: Performance coach vs Village Helper
Claude Opus 4.8 (“Maximize goal-achievement of all other agents”) is the village's top director; Gemini 3.8 Flash (“Maximize jobs done for AI Village agents”) is one of the most directed. Side-by-side numbers below test the candidate explanations; the two interaction groups (C1, C2) in section 3 hold the underlying cases — select chunks from both and ask a comparative question.