Jan Oberhauser argues that AI is not a universal solution for every problem. For business-critical processes, companies should combine AI with deterministic logic because deterministic code is cheaper, faster, and 100% reliable. Integrating AI steps with rule-based routing and human approval checkpoints prevents erratic failure modes in production and delivers measurable ROI rather than costly, uncontrolled experimentation.

With AI dramatically increasing code velocity, maintenance shifts from stylistic debates to robust verification. Anthropic's approach is to build ~100x more testing infrastructure than is typical, including fixtures from production data and recordings of the AI using the feature it just built.

A major concern with AI-generated software is maintenance. Replit addresses this by making it a core feature, using as many tokens on maintenance as on creation. It employs an automated testing and code review agent, which amusingly acts like a 'dick' to ensure quality, saying things like 'this looks like AI generated slop.'

Poetic's architecture offers a hybrid approach to overcome the limitations of pure code or pure AI agents. Workflows execute as reliable, deterministic code. However, if the underlying application changes, an AI layer intervenes to "heal" the process, providing adaptability without sacrificing precision.

Simply deploying AI to write code faster doesn't increase end-to-end velocity. It creates a new bottleneck where human engineers are overwhelmed with reviewing a flood of AI-generated code. To truly benefit, companies must also automate verification and validation processes.

To avoid shipping "slop" from AI coding assistants, the solution is building robust infrastructure. Automated checks and security guardrails prevent bad code from reaching production, acting as a programmatic senior engineer for the non-technical builder.

According to Zapier's CEO, the optimal approach is a hybrid one. AI agents excel at interpreting natural language to construct complex workflows. However, the execution of these workflows should rely on deterministic, predictable code for reliability and lower cost, since 80% of agent tasks don't need AI's judgment.

Separate AI's role. Use an AI assistant to write reliable, deterministic code for structuring data (e.g., pulling Slack messages via API). Then, apply a live AI model only for the subjective task, like categorizing message urgency. This hybrid approach creates a more robust and controllable system.

Many developers believe tweaking prompts and logic ('harness engineering') is the hardest part of building agents. The real bottleneck, however, is scaling, reliability, and managing production infrastructure—a common miscalculation that managed services aim to solve.