Fifteen Months of Fintech with Agents: The Quality Curve Across the Project's History

Alexander, a remote developer at a fintech startup, has been running a closed fintech product with AI agents writing the code since May 2025, with a second payment project added in February 2026. He observed that AI agents produce a distinct class of errors, such as error masking, fallback returns, and empty struct returns, which standard linters don't detect. He wrote 50 custom analyzers within the project, then consolidated them into an open-source tool called glint, which now has 130 rules. The quality curve of the project shows plateaus and drops, each drop corresponding to specific events like the January cleanup (removing 140K lines of dead code) and the April god-object analysis. The developer measured the density of severe findings per thousand non-test lines of Go code, and the curve fell below his own hand-written code quality in April 2026, becoming six times cleaner. He also ran a control group project (without the lint circuit) that showed worsening quality, and a 2024 crypto bot he wrote without agents had 2.32 severe findings per thousand lines, including unsafe type assertions and money in float JSON contracts. The system works as a whole: lint rules catch defects before commits, and the first to know about a bug is the checker, not the client. However, the lint does not catch everything: a recent bug where money was deducted but displayed as not deducted slipped through three layers of protection. The developer concludes that quality is now a property of the circuit in which code is written, not of the writer.