This is the authors' abstract. We don't add key points for this paper.
Agentis v0.8 is the operational-process layer of the SCF × DVE × VEROK research program. It treats closed commercial AI systems as capabilities placed in defined seats of a workflow and builds an observation layer out of process structure, without access to model internals. This report presents the 32-day operational record of Agentis v0.8 (2026-10-13 to 2026-11-14) and an independent read-only audit of that record. Two commercial agents worked in the governed setup: Claude in the Synthesis seat and Codex in the Construction-Unit seat. Over the period the system produced 516 mailbox documents (203 work tickets, 231 reports, 55 proposals), 428 author confirmations, 10,660 self-reported metadata events, 22,867 independent filesystem-watcher events and 2,241 completed rule-evaluation cycles. On 2026-11-14, Google Antigravity (Gemini 4 Hydrogen and Gemini 4 Argon) audited this record through three read-only specialist subagents. The audit reports three findings: Observability held; the setup plane absorbed the effort. A single write gate, an independent filesystem watcher and a non-AI rule engine with veto/hold authority kept the work observable and blocked silent self-certification. However, 90.06% of all events (9,601 of 10,660) belonged to maintaining the observability tooling itself, and no operational-plane task was executed. Agent errors rose again as context accumulated. The agents recorded 84 event corrections and 19 blocked executions. After the correction rate had fallen to 0.08% in week 3, it rose to 0.75% in the final twelve days, when the agents' append-only context files had reached roughly 266–692 KB. Self-reported test passes repeatedly masked failures that only appeared against the real system. Full-history checks scaled linearly and overflowed the cycle. As the three core ledgers grew 2.95-fold in two weeks, the median rule-engine cycle grew 3.51-fold (16.66 s to 58.46 s), and in the final days the 95th-percentile cycle reached 308 s and the longest completed cycle 404 s, beyond the 300 s scheduling window.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Thanh Nam Nguyen (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: