Running an Autonomous AI Agent on Quantitative Trading Research: Protocol, Supervision and Failure Modes

In October 2026 a private research project on equity strategies was run almost entirely by an AI coding agent: 95 version-control commits in about 46 hours, from a one-page rule file written by the human and a standing instruction to work autonomously. This paper documents the working method and what went wrong. It describes the rule set and the reason for each rule, the supervision pattern (goals and hard rules from the human, autonomous work, periodic independent re-reading of the rules), the pre-registration convention with dated addenda, parallel sub-agents with blind labelling, and the disclosure of post-cutoff leaks. It then lists failure modes specific to language-model agents: look-ahead through the model's own knowledge, confident but contradictory documents, prorated numbers presented as measured ones, numbers that drift after data reloads, and silent failures of scripted edits. For every rule we give one anonymised incident from the project history: what went wrong, how it was caught, the fix and the regression test. We also state the access rules (robots.txt, per-host day lock after HTTP 429, neutral User-Agent, no captcha or paywall circumvention, read-only use of personal data feeds) and the final step of the method, which is to turn the repositories into papers using a critic-and-reviewer loop. The evidence is one project and is self-reported; the paper is a field report and a checklist, not a controlled study.Read next: The guard rails the agent worked under are the leak-resistant point-in-time backtesting. The agent's final verdict was a pre-registered out-of-sample test. Most of its findings were negative results on event-driven signals.This preprint is part of an eight-paper series on retail equity research from public data by Christopher Helm, covering German and Nordic small caps, backtesting methodology, data sources and AI-assisted research. All papers in the series:How Wise Is the Crowd? Timing and Selection Skill of Retail Social-Trading PortfoliosA Leak-Resistant Point-in-Time Backtesting Protocol and Engine for Retail Equity StrategiesThree Pre-Registered Bets: A One-Shot Out-of-Sample Test of Retail Strategy CandidatesWhat Did Not Work: Negative Results on Event-Driven Signals in German and Nordic Small CapsFree, Point-in-Time Public Data for German and Nordic Small-Cap Research: Sources, Coverage and PitfallsRequirements for an Autonomous Retail Order Engine Running Unattended for a DecadeFrom Pre-Announcement to Verified Facts: Design of a Primary-Source Report-Analysis Funnel for Small-Cap EquitiesChristopher Helm is an entrepreneur, consultant, and researcher. More research and articles on his blog.