The Coming Loop: Agentic Engineering and Code Quality

Armin Ronacher, a prominent software engineer, reflects on the evolution of coding agents and agentic engineering patterns. He observes that developers are increasingly building systems with two distinct loops: an inner agent loop where the model calls tools and iteratively solves problems, and an outer harness loop that manages tasks beyond the agent's natural stopping point. This outer loop pattern—queuing work, attempting execution, and deciding whether to continue, modify context, or restart—has become prevalent in recent agentic engineering discussions. Ronacher notes this approach appears in various implementations, including Anthropic's Claude Code and other platforms. However, he expresses concerns about code quality from hands-off harnesses. He states that models tend to produce overly defensive code with excessive complexity, poor abstractions, and code duplication. The fundamental issue is that models avoid strong invariants and add local error handling instead of making bad states impossible. Ronacher observes little progress in improving these tendencies and suspects the direction may have worsened, with extended uninterrupted model operation (30+ minutes) producing worse outcomes than earlier human-in-the-loop approaches. He cites research noting models are 'mortally terrified of exceptions,' leading to defensive patterns inappropriate for systems with important invariants like data formats or core infrastructure. His conclusion suggests that extending hands-off agent loops without human oversight may amplify these code quality problems.