XNews

Engineering practices for agents and AI systems

Engineering practices for agents and AI systems

Engineering practices for agentic systems and AI code workflows. - Two-version approach: prototype with AI, then rewrite for production - Prototype confirms requirements and technical landmines cheaply - Production version applies design, CI, reviews and maintainability - LangChain: Deep Agents evaluation requires custom tests per case - Testing modes: single-step, full turn, multiple turns - Test objects: trajectories, final response, and state artifacts - Single-step tests cost-efficient; full-run tests validate end-to-end behavior - Environment isolation (containers, recorded HTTP) crucial for reproducibility

Key insights

Two-pass development reduces risk and accelerates discovery

Practitioners recommend building a disposable prototype using AI to validate requirements and surface technical unknowns, then re-implementing a production-grade version with proper architecture, CI, and code review. The prototype minimizes wasted engineering time and clarifies what must be designed robustly in the second pass.

Deep Agents need tailored, trajectory-aware evaluation

LangChain's operational guidance shows that evaluating deep, multi-step agents requires custom assertion logic per test case and a mix of single-step, full-turn and multi-turn tests. Key artifacts include the tool call trajectory, the agent's final response, and generated files/state. Sandboxed reproducible environments and HTTP recording speed up reliable tests.

Source posts