As of October 28, 2026, the job market shows a high demand for AI agents, with 1,458 job postings mentioning the term on hh.ru. Other common requirements include LLMs in 1,261 listings, RAG in 520, and MCP in 245. In contrast, prompt caching and token management appear in only 13 and 19 vacancies, respectively.
AI engineering focuses on building systems around pre-trained models rather than training them. Engineers provide data, tools, and constraints to the model, which acts as a stateless service. According to a 2026 Stack Overflow survey, 73% of developers use AI assistants daily, with Claude Code and GitHub Copilot being the most frequently cited tools.
Models operate within a context window, which includes all instructions and history sent with every request. For Claude models, this window is typically 1 million tokens. Because models are stateless, the system must resend the entire history at each step, leading to significant token consumption. Using prompt caching can reduce costs by 5–10% for repetitive inputs.
The cost structure for Claude Sonnet 5.5, as of October 28, 2026, is $2 per million input tokens and $10 per million output tokens. Since output tokens are five times more expensive, managing the agent's cycle and context is critical for budget control. Developers use hooks and sub-agents to optimize these processes and prevent unnecessary token usage.
The Model Context Protocol (MCP) serves as a standard for connecting AI applications to external systems, similar to USB-C. It allows for the integration of tools and data sources like GitHub or databases. The author notes that the standard is evolving rapidly, with the reference point being the version released on July 28, 2026.
