I Had Already Built Three Agentic Loops Without Naming Them

Описание и применение

Автор описал, как он сформировал три вложенных цикла обратной связи для автономного кода‑генератора: план выполнения, контрольные точки и TDD с архитектурными тестами, что повысило автономию агента и снизило ручные проверки.

Задача

How to give an AI coding agent enough autonomy to implement a feature while ensuring it stays on target, respects architecture, and delivers observable progress?

Результат

By structuring the workflow into three nested loops, the author achieved higher agent autonomy, reduced manual checkpoints, and ensured that each stage of development—planning, milestone completion, and behavior/architecture compliance—was independently validated.

Рабочий процесс

• Define an execution plan that includes purpose, context, desired behavior, constraints, milestones, validation strategy, and acceptance criteria.
• The agent executes the plan, updates it as it discovers new information, and continues until the final acceptance criteria are met.
• Within the plan, the agent follows milestones, each with its own scope, validation commands, expected result, and acceptance criteria.
• At each milestone, the agent runs TDD cycles: write failing test, implement minimal code, refactor if needed, and repeat until tests pass.
• During TDD, architectural tests (e.g., HexagonalArchTest) run to enforce design boundaries; failures trigger design corrections before proceeding.
• After a few TDD cycles, the agent validates the feature through its public path to prevent drift into implementation details.
• The agent continues until all loops report success; the author then reviews the pull request for final approval.

Доказательства

The outermost layer is the execution plan.

Источник

Inside the milestones is the smallest and most active layer.

Источник

The agent does not become more autonomous because I stopped caring about the process.

Источник