Prompt vs Loop vs Graph Engineering: три уровня контроля в AI-разработке

MarkTechPost, on July 29, 2026, broke down three trendy AI engineering terms — prompt engineering, loop engineering, and graph engineering — as three nested levels of control, rather than competing approaches.

Three levels of control

Prompt engineering, loop engineering, and graph engineering are not rival techniques but different units of control, stacked on top of each other. A prompt controls a single model response, a loop controls the behavior of a single agent, and a graph controls the organization of multiple agents. Each upper layer preserves the one below it: the prompt doesn't disappear when a loop is built around it — it simply stops being something typed by hand.

The terms emerged in practice at different times. Prompt engineering is well-established. Loop engineering entered the vocabulary in late 2025 and dominated developer discussions until June 2026. Graph engineering followed about six weeks later and is currently the least settled of the three.

What the stack is made of

The full progression consists of five stages: prompt → context → harness → loop → graph. Each was named in practice before it made it into vendor documentation.

• Prompt engineering — writing instructions for a single call; Anthropic recommends splitting the system prompt into sections (background, instructions, tool use, output description) using XML tags or Markdown headings
• Context engineering — per Anthropic's description, a natural extension: the question shifts from "which words to choose" to "what configuration of tokens should even be in the window"
• Harness engineering — the environment of a single agent: files, tools, memory, feedback
• Loop engineering — one floor up: how the system repeatedly observes, acts, verifies, and recovers
• Graph engineering — orchestration of multiple agents based on a graph

A June 2026 arXiv paper on agentic AI in construction engineering (the Buildrix project) explicitly lays out the same chain: prompt, then context, then harness, then loop.

Where a single prompt breaks down

Prompt engineering assumes a human is present at every iteration: the prompt is written, the model responds, the output is evaluated, the prompt is rewritten. It's precisely this assumption that breaks. High volume, multi-step tasks, no human available for evaluation, results that get automatically fed into the next step — any one of these conditions is enough for a single prompt to stop being sufficient.

The prompt itself doesn't get worse in this process — the surrounding conditions change. And within the upper layers, prompt engineering doesn't disappear: according to Anthropic's report on multi-agent research, prompts were the primary lever for fixing coordination failures. Early versions spawned up to 50 subagents for simple queries, and this was fixed with prompting, not topology.

"Prompt engineering was our most reliable lever for fixing coordination failures,"

Anthropic's research report on multi-agent systems states.

What loop engineering is

Loop engineering treats a coding agent as a brute-force tool for searching for solutions, with mastery meaning designing the goal, the tools, and the loop itself — not just the prompt. The term went mainstream in June 2026 after a widely shared post whose author called on engineers to stop "prompting" coding agents and start designing loops that prompt them instead. The same week, the Claude Code team at Anthropic described the same shift from the stage.

The most detailed public breakdown identifies five primitives and a sixth element binding them together: automations (a schedule or event triggering discovery and triage without supervision), worktrees (isolation so parallel agents don't edit the same files), and others.

What this means

The three terms describe not a trend but three distinct units of control — model response, agent loop, and agent organization. Understanding which layer a task sits at determines exactly what to design: the words of a prompt, a feedback loop, or a coordination graph.

FAQ

How does loop engineering differ from prompt engineering?

A prompt controls a single model response, while a loop controls an agent's behavioral cycle: how it observes, acts, verifies, and recovers. A loop doesn't replace the prompt — it's built on top of it.

When did the term loop engineering appear?

According to MarkTechPost, loop engineering entered the lexicon in late 2025 and dominated developer discussions until June 2026, when it went mainstream after a widely shared post.

What is graph engineering?

Graph engineering is the newest of the three terms, appearing about six weeks after loop engineering. It refers to orchestrating multiple agents based on a graph; the term's origin hasn't fully settled yet and overlaps with the older meaning from knowledge graph.