The Limits of Prompting: ArchitectingTrustworthy Coding Agents

LLMs can summarize, autocomplete, and comment on code, but when applied to tasks like code review, they often fall apart. We’ve run agents across real PRs in production environments and observed consistent failure modes: overconfidence, repetition, and lack of judgment. This talk examines the architectural reasons behind that breakdown—why transformer-based models struggle with intent, structure, and impact when reviewing code - and how reasoning phases, memory, and reflection improve both precision and trust. Using real-world examples from Baz’s code review agent, we’ll show why prompt engineering isn’t enough, and what it takes to move from generic pattern-matching to useful, reliable feedback in large, evolving codebases. Expect concrete takeaways for anyone building or evaluating LLM-based developer tools.