Agentic AI sucks (the life out of me) (whatblog.ghost.io)

A veteran developer recounts building an “agentic” system on top of OpenAI’s LLM services and finding the experience dispiriting: the LLM is not just a component but the unpredictable core. Although the surrounding infrastructure—workflows, agents and tools—was straightforward to build and the system “works,” the model’s stochastic behavior, long multi-turn executions, and opaque internal state make outcomes unreliable, slow, and effectively untestable without full end-to-end runs. The author describes a workflow where inputs only probabilistically map to outputs, fixes often mean re-running until it “eventually” does the right thing, and there’s a painful temporal gap between action and result. For the AI/ML community this is a practical wake-up call: agentic systems expose gaps in reproducibility, observability, and developer ergonomics. Technical implications include urgent needs for deterministic or versioned execution modes, better tooling for unit-testable agent steps, richer provenance and logging for multi-turn context, and performance/latency improvements so rapid iteration is possible. Beyond engineering, there’s a cultural cost—loss of developer joy and control—highlighting that handing core creative work to brittle models risks degrading product vision. The note argues we must prioritize debuggability and predictable behavior if agentic AI is to be a reliable part of software engineering.