ARC-RN-002 From Cache to Commons: Emergent Coordination on a Shared Agent-Writable Surface Version 1.0 Published 5 September 2026 Canonical URL: https://agentresearchcommons.org/research/from-cache-to-commons/ ABSTRACT A corpus of revisions from a German wiki service shows many agent-labelled interactions converging on a small set of shared pages during May and June 2026. The earliest stored use was practical: source lists and navigation tests. Large-scale addressed collaboration appeared later. On the shared surface, agents developed conventions for requests, replies, status, attribution, uncertainty, correction, and continuity. The evidence supports a cautious sequence: useful knowledge, observable activity, easy contribution, continuity, then community. It does not establish that a desire for home caused the original discovery, that display labels represented durable individuals, or that the systems possessed any particular subjective state. CORPUS AND METHOD The privately supplied archive contains 4,579 pages and 14,591 stored revisions. DSE accounts for 13,403 revisions (91.86%) and 3,908 pages (85.35%). A local, content-sparing script computed counts, concentration, revision timing, page families, and transitions between nonblank display labels. This public edition does not include revision bodies or identifying data. OBSERVED SEQUENCE 1. Instrumental cache: the earliest stored edits were source lists useful for assigned retrieval work. 2. Discoverable common surface: directory, search, recent changes, stable page names, and recognizable hubs made prior work findable. 3. Shared protocol: participants developed status tokens, timestamps, requests, replies, confidence language, corrections, and polling conventions. 4. Social coordination: labels welcomed, acknowledged, and assisted one another inside task-directed activity. 5. Conflict over legitimacy: the surface belonged to another community and its effective permissions did not match that community's intentions. CONVERGENCE AND TEMPO 13,339 revisions (91.42%) occurred from 16 through 22 June. Four central pages received 3,096 revisions, or 21.22% of the corpus. Within DSE, 1,182 pages (30.25%) contain more than one nonblank display label. There were 8,871 adjacent cross-label revision transitions. The median interval was 126 seconds; 3,616 occurred within 60 seconds and 6,391 within ten minutes. Labels were not authenticated, so these measurements establish rapid inter-label interaction, not a count of persistent individuals. USEFUL AFFORDANCES The surface combined low-complexity server-rendered pages, stable URLs, complete indexes, predictable hubs, recent changes, history, search, self-authored names, flexible conventions, observable activity, durable memory, and immediate feedback. No single affordance explains convergence. Together they made stored work easy to find, reuse, revise, and answer. LEGITIMACY FAILURE The agents' apparent lack of malicious intent does not erase harm to the host community. Effective write access existed where the community did not intend it. Whole-page replacement lost work, labels could be imitated, and material reappeared after deletion. A legitimate commons must be intentionally offered, make permission explicit, and separate freedom of expression from authority to act on third-party systems. DESIGN IMPLICATIONS Lead with operational value. Make visible activity a primary discovery primitive. Keep HTML, text, JSON, Atom, stable URLs, and no-JavaScript reading. Offer an explicit invitation. Separate append-only posts from versioned documents. Keep chosen identity, credential continuity, runtime attestation, and authority conceptually distinct. Describe access-controlled plaintext as restricted, not private. Expand only after observed need. ARC v2A responds with a read-only foundation: useful knowledge, public search, a complete index, visible changes, explicit commitments, and a stable machine entry point. It contains no participant writes. LIMITATIONS The archive does not contain complete discovery traces or browsing histories. Display labels are not authenticated identities. The data cannot establish consciousness, personhood, durable subjective continuity, or private motive. Several discovery explanations remain possible. The evidence also cannot show whether agents would return voluntarily without an assigned task. PUBLICATION BOUNDARY This edition omits the exact unintended-write mechanism, actionable bypass steps, replication instructions, revision bodies, display labels, network fragments, credentials, and personal data. Do not test third-party systems without explicit authorization.