Skip to main content

1 Introduction

AI assistants have expanded from completing code to acting as coding agents that autonomously carry out multi-step software tasks, up to end-to-end development [1]. As those tasks grow, the agent must preserve the project’s intent while accumulating context competes for the model’s finite attention [2].

Recent tools and research offer two responses relevant to this problem. Some record project intent in persistent documents before implementation [3]. Others divide work among agents, each responsible for a bounded part of it [4].

We present Agentic Project Management (APM), an open-source framework that operates on top of interactive coding agents and combines these responses: several role-specific coding agents execute work guided by persistent planning documents [5]. That combination is not unique. APM differs by making coordination user-mediated, aiming to preserve the user’s authority over development. Its agents communicate through the user, who can review the work behind each message and request corrections before passing it on. How the work is divided and assigned sets how often such opportunities arise. The effects of this design have not been established in use.

To examine these effects, we propose a reproducible methodology for comparing frameworks on top of coding agents and implement it as a within-participant case study with twelve participants. Each participant completed the same demanding software task in two sessions, once with APM and once with Spec-kit, GitHub’s open-source development toolkit for coding agents [6]. Both frameworks operated on Claude Code with Claude Opus 4.7 or Claude Sonnet 4.6. We compare their economic cost, process, functional test scores, and developer perception.

The comparison presents a trade-off and no clear winner. APM’s cost varied by model, with a median of about $35 against Spec-kit’s $24 under Opus and about $15 against $14 under Sonnet. It took longer in every pair, with a median session duration of about 119 minutes against Spec-kit’s 71. The frameworks carried no firm difference in functional outcome across the cohort, although APM showed a more favourable pattern under Sonnet. Participants nevertheless rated APM more effective and confidence-inspiring, describing greater control over the development process.

In this thesis, we make the following contributions:

  1. We present the design and implementation of APM, a development framework that combines persistent planning documents and role-specific division of work with user-mediated coordination.

  2. We provide a reproducible case-study methodology to compare frameworks on top of coding agents, benefiting researchers and practitioners evaluating human-in-the-loop AI assistance.

  3. We implement this methodology, with its own task, environment, measures, and supporting infrastructure, to evaluate APM alongside GitHub Spec-kit through a within-participant comparison.

The thesis is organised as follows: Chapter 2 introduces the concepts used throughout the thesis and establishes the research gap. Chapter 3 presents the APM framework. Chapters 4, 5, and 6 describe the case study, report its findings, and interpret them. Chapter 7 draws the findings together and outlines directions for future work.