Loop Engineering: Designing Systems That Run Coding Agents

Alexander Polomodov and Maxim Smirnov describe Loop Engineering as the shift from one model call to a system that discovers work, launches agents, and verifies outcomes. HuaShu’s playbook maps current practices, but it is a working note rather than peer-reviewed evidence of effectiveness.

A loop combines discovery, isolated handoff, verification, state outside the conversation, and scheduling. Worktrees, skills, connectors, sub-agents, and memory support these functions; a timer without fresh input and feedback merely repeats the prompt and amplifies mistakes.

The loop closes through an independent evaluator with clean context and authority to reject results. It tests observable behavior by running tests, opening the interface, calling APIs, and checking data. Green tests and an LLM judge are insufficient, so deterministic gates and human acceptance remain necessary.

Autonomy creates verification debt, lost understanding, and token waste. A first loop needs one queue, an isolated worktree, external state, and independent evaluation. Unattended use also needs limits, prompt-injection defenses, least privilege, audit, rollback, and a kill switch; expansion should follow measured quality, cost, and review effort.