
November 23–27, 2026
Principal Dev Engineer
Capital One
I gave AI coding agents one architecture rule, “services never read another team’s database tables”, and tried to get them to break it. With no guardrail, they violated it in 38 of 50 runs: agents don’t ignore your architecture decisions, they never see them. Once I added an interception check at tool-call time using gatekeeper-mcp, an open-source Model Context Protocol (MCP) server, direct table reads stopped. Then it got interesting: my rule said tables, and the agents found the gaps, routing around the check through shared reporting views or by modifying schema paths. Six of 14 bypass attempts got through. Those are the dangerous ones, because the diff looks clean: AI code is plausible by design. That is what I call Reviewer Atrophy: plausible-looking code lowers reviewer alertness faster than velocity rises. Sonar’s 2026 survey of 1,100 developers puts numbers on it: 96% of developers don’t fully trust AI-generated code, yet only 48% always verify it before committing. By the time a reviewer spots the violation, the agent has built three more features on top of it. ArchUnit and OPA typically run in CI or pre-commit, after the agent has built around the mistake. Here, an architecture decision record (ADR) becomes a rule the agent hits before a pull request exists. I’ll live-demo an agent getting blocked and fixing itself, then show what stopped the agents, what didn’t, which ADRs are worth automating, and how to write rules that survive an agent hunting for gaps. For engineers and tech leads using AI coding assistants like Copilot, Claude Code or Cursor.
See every session across the AI Agents Conference November 2026 line-up.
View Agenda