7 Conclusion
This thesis presented APM v1, a ground-up redesign of the open-source APM framework. It operates on top of coding agents, combining persistent planning documents and role-specific division of work with user-mediated coordination. It also proposed a reproducible case-study methodology for comparing frameworks of this class, and implemented it to compare APM v1 against Spec-kit through a within-participant case study with twelve participants, covering economic, process, functional, and perception outcomes.
Across the cohort, the frameworks carried no firm difference in functional outcome. That comparability is what a grader reading only the running program’s responses can see. Section 6.3 bounds what it reaches.
APM cost more, and its several conversations account for that. Each accumulates its own context, builds its own cache, and produces its own output. At session scope their costs add together. Under Opus the whole task fit within one context window limit and no conversation reset, leaving nothing to offset that aggregate against Spec-kit’s single conversation. Under Sonnet the same task repeatedly pushed Spec-kit towards Claude Code’s auto-compaction threshold, and it reread a large accumulated cache before each reset. APM instead kept its conversations’ context smaller and compacted less, bringing the two frameworks’ costs close, as Section 6.2.1 explains.
The APM session lasted longer in every pair and had more turn prompts in eleven of the twelve, with its additional interaction and time falling together in implementation. APM coordinates several agent conversations through the user by design, and both follow from it. Each Task Prompt and Task Report passes through the user and opens a turn in the conversation receiving it. More turn prompts open more intervals of agent work, and turn time made up most of that phase’s longer duration, as Section 6.2.2 reports.
Against those additional process and economic demands, participants rated APM higher on effectiveness and confidence, and described greater control over the development process. Our interaction measurements show the user-mediated coordination operating. They leave open whether participants used it to review code, prevent defects, or improve the resulting software. The findings remain bounded by the factors Section 6.7 discusses. The methodology is meant to apply to other frameworks of this class, with a task, environment, and tooling suited to each comparison.
7.1 Future Work
This thesis concentrated on designing and building APM v1, and on designing the evaluation methodology and building its infrastructure. Several directions that follow from that work remain open, and we outline them for researchers who wish to pursue them.
Task size and work decomposition. It remains open how task size and demand shift the value of multi-agent systems and their decomposition of work. Future evaluations of frameworks on top of coding agents could vary them, comparing single-agent continuous workflows with multi-agent decomposition across models of differing capability and context limits. That comparison would show how the fit between task demand and context limits shifts the trade-off between an accumulating conversation and distributed agent contexts. Such evaluations would also need an unguided agent baseline. A less capable model may require additional guidance to succeed, whereas a capable model may manage alone, rendering framework structure pure overhead. Rotating these factors against one another would isolate where such decomposition provides genuine economic and functional value. Monitoring how users size and dispatch APM Tasks would also show how often exchanges pass through the user, how often users take these opportunities to review the work, and what that costs.
Maintainability of produced software. Software is read, changed, and repaired long after it is delivered. Deterministic tests verify immediate correctness, but leave architectural degradation and downstream maintainability unmeasured. Rapid code generation risks producing code bloat and superficial solutions that satisfy immediate checks yet complicate later maintenance. A longitudinal evaluation following software produced by APM alongside other frameworks of this class would track defect correction, code churn, and lifecycle cost. Two properties of APM’s design bear directly on this: focused Workers operating on scoped contexts to resist context decay, and the user-mediated coordination mechanism. Such an evaluation would test whether these mechanisms protect against software degradation, repaying APM’s upfront coordination overhead with lower maintenance costs later.
Developer comprehension. Beyond coordination mechanics, user mediation raises questions about developer comprehension and authority. The turn-prompt buckets record the form of user input but not its content; classifying that content across the session, and for each APM role, would show if, and how, participants exercised their authority at each friction point. Our case study measured self-reported confidence and perceived control, but left technical understanding unmeasured. That gap leaves open whether APM fosters genuine program comprehension or an illusion of control. Such an illusion could add to the cognitive debt that builds as developers offload mental effort to AI assistants [58]. Future studies could adapt empirical comprehension methods, using structured mental representation tasks [59] to test developers objectively on program structure, data flow, and state transitions depending on the task at hand. Evaluating comprehension accuracy alongside subjective ratings would determine whether the framework’s design preserves usable developer authority or merely inflates perceived confidence.
User mediation against autonomous coordination. Finally, it remains open what user mediation itself contributes. APM’s user-mediated coordination could be compared directly with an autonomous variant whose agents communicate without the user, such as the APM Auto design described in Section 3.7. A controlled comparison would isolate the contribution of human oversight from the coordination of its agents, measuring whether removing the mechanism accelerates delivery at the expense of autonomous drift, undetected error propagation, and downstream maintenance debt across complex dependency graphs.