Agentic Coding and reality: why prompting is not enough for real software

The tech industry loves extremes. A year ago the trend was 'prompt engineering will replace programmers.' Today it is the hype around 'Dark Software Factories' — fully autonomous loops of AI agents building entire products without human intervention.

In his talk 'Graph Hysteria & Agentic Reality', Ilya Klimov takes a sober, critical look at these promises. While AI agents are extraordinarily capable, pretending that software development can be reduced to a black-box graph of LLM nodes ignores hard engineering realities: model drift, non-deterministic errors, and cascading failure states.

Agentic Coding is not about writing clever prompt templates. As Ilya notes: 'A prompt is the weakest possible contract in software. Real engineering begins when you constrain the model with strict schemas, atomic tool APIs, state persistence, and human checkpoints.'

When building production agent workflows — whether using LangGraph, Vercel AI SDK, or custom MCP (Model Context Protocol) servers — the secret to reliability is small, scoped nodes. A monolithic prompt trying to handle authorization, data transformation, and API calls will fail 20% of the time. A multi-step flow where each node performs one validated action will succeed predictably.

From my experience shipping FinTech and SaaS web applications: human-in-the-loop design is not a compromise — it is a feature. The best architectures use AI agents to automate 90% of the research, drafting, and test writing, while reserving critical state changes (money transactions, database writes, user messaging) for validated human confirmation.

If you are designing agentic software today: start simple. Don't build an infinitely complex graph of 50 agents on day one. Build typed tools, enforce strict schema validation, set hard execution limits, and keep the human in the loop where the stakes are real.