Cellular Mode: A Cell-Based and Verifiable Method for Agent-Assisted Software Engineering

Abstract

This preprint describes Cellular Mode v1.0.0, an open-source method and software architecture for bounded, resumable, auditable, and verifiable agent-assisted software engineering. It has not undergone peer review. Abstract Large language models (LLMs) and agentic coding systems can now inspect repositories, modify multiple files, execute tests, and participate in long-running implementation cycles, yet the organization of sustained human-agent development remains comparatively under-specified. This paper presents Cellular Mode, an open-source method and software architecture for bounded, resumable, auditable, and verifiable agent-assisted software engineering. The method treats the cell as the primary unit of work: a bounded task with explicit scope, state, acceptance criteria, recorded decisions, and a concrete next action. Persistent project state is maintained outside conversational context through an append-only event log and reconstructible projections, enabling work to be paused and resumed without requiring continuity of model memory. A separate engineering layer introduces explicit epistemic states, mechanical gates, human authorization, and Article 8 final-state verification, which binds verification evidence to the exact Git tree later committed. The adoption layer includes a Prompt Builder and Bootstrap mechanism for new and existing repositories, while optional modules provide adaptive interaction, read-only observation, and polyglot plugin capabilities through the Universal Plugin Protocol. Version 1.0.0 was evaluated through deterministic tests, independent continuous integration on Node.js 22 and 24, protocol-conformance tests across multiple executed language implementations, optional-component removal rehearsals, and two structured adoption walkthroughs conducted by an AI agent acting as a technically competent newcomer. These evaluations support claims about implementation behavior, state integrity, and workflow coherence; they do not constitute a human usability study. The software is released under the Apache License 2.0 and archived on Zenodo as a version-specific research software artifact.