Three layers, not competitors

Prompt, loop and graph engineering are stacked “units of control”, not competing methods. A prompt steers one response; a loop steers one agent's repetition cycle (goal, tools, repetition, stop condition); a graph steers the organisation of multiple agents through nodes and edges. Each layer keeps the layers beneath it intact.

The terminology is young: loop engineering broke through in June 2026, graph engineering as a term in July 2026 — helped by tooling such as LangGraph (StateGraph, conditional edges) and the agent features in Claude Code and Codex.

What makes a loop good

Six elements recur in descriptions of loop engineering: scheduled automations, worktrees for parallel isolation, reusable skills, plugins/connectors (often via MCP), sub-agents with a maker/checker split, and persistent state outside the conversation.

The main pitfall is measurability: a loop that cannot mechanically distinguish “done” from “stuck” does not fail loudly — it keeps spending tokens. Verifiable success criteria (tests, schemas, rubrics, or a second model as judge) are a precondition, not a luxury.

What a graph adds

Graph engineering describes multi-agent systems as a topology: node functions receive state and return partial updates; context does not cross a node boundary unless an edge carries it explicitly. Production systems distinguish a stable “org graph” (fixed roles) from an ephemeral “work graph” (the current task, with splitting, merging and cancelling).

Caveat: for most solo projects a graph is overkill. The common selection framework is a series of questions — does human review precede every action? can completion be verified non-manually? does the task fit one agent's context? must independent branches run simultaneously? — where the first “no” tells you which layer you need.

For the indie developer

Practically: start with a good prompt, move to a loop only once you have verifiable success criteria, and to a graph only when independent subtasks genuinely must run in parallel. For the rules of reliable AI coding that precede the loop layer, also read our explainer of Karpathy's four rules; daily agent-tooling signals are on AI Radar.

Sources: MarkTechPost, July 29, 2026; background: LangGraph documentation and the agent features of Claude Code/Codex. Verified July 31, 2026.