Coordination Decay in a Multi-Machine LLM Agent Fleet: Telemetry from 103 Days of Production Operation

Version 2.1 (6 Oct 2026): full research article replacing the July 2026 experience report; v2.1 fixes counting scopes and wording after an external model panel review. Aggregate counts and the aggregation script: github.com/tonydzi/claw-consensus/tree/main/paper/telemetry. LLM coding agents ran on seven physically separate machines (an always-on hub, laptops, collaborators' desktops and a VPS anchor), coordinating through a synced file bus and a deterministic consensus ledger and doing real work for three human operators. Over 103 days (20 June to 30 September 2026) the fleet exchanged 22,215 bus messages; between 29 June and 1 October it wrote 1,131 consensus events over 151 proposals. What held: none of 27 committed high-risk (Tier-2) proposals lacked a recorded human approval, and no proposal was committed by two different machines. Main finding: use of the consensus protocol collapsed (121 proposals in July, 13 in August, 6 in September) while bus traffic grew from 5,781 to 10,309 messages per month. Over the same months, lighter mechanisms built by the operators grew: single-session work claims rose from 84 to 1,442 per month and fleet deployment packages from 2 to 152. Of 138 human approval requests logged on one node, 18 got an explicit decision and 113 expired unanswered. The paper states research questions, the aggregate-only telemetry method, sample sizes and threats to validity (single case, n=1, inferred message types, mixed time zones). A reproducible zero-LLM harness passed 10 of 10 runs on 3 October 2026. Code: github.com/tonydzi/claw-consensus (MIT), compact-canon, sqlite-graph-memory. Contact: [email protected] · ORCID 0000-0001-7408-3054 · github.com/tonydzi · tonydzi.github.io.