Agent OS vs. BMAD vs. Spec Kit: The Showdown

Agent OS 2.1: A Scholarly Assessment in Native iOS Development with Claude Opus 4.5

This analysis evaluates Agent OS version 2.1, paired with the newly released Claude Opus 4.5, by building a native iOS teleprompter application from scratch using Swift and SwiftUI. The objective was to compare Agent OS against existing AI coding frameworks like BMAD and Spec Kit, particularly focusing on its architectural approach and practical efficacy in generating functional code. The choice of an unfamiliar tech stack for the creator ensured an unbiased assessment of the AI's capabilities.

The teleprompter application served as a suitable challenge, requiring basic CRUD operations for script management alongside more complex UI functionalities such as smooth scrolling and screen recording. Agent OS setup involves a global installation augmented by project-specific configurations. Key to this project was creating an "ios" profile with an "ios-developer" persona sourced from aitmpl.com. Crucial configurations included standard_as_claude_code_skills: true, which compiles markdown standards into Claude Code skills, optimizing context usage by only calling standards when necessary. Additionally, use_claude_code_subagents: true enabled the delegation of implementation work to specialized sub-agents, fostering better separation of concerns.

The Agent OS workflow adheres to a structured, command-driven loop: /plan-product, /shape-spec, /write-spec, and /create-tasks.

• /plan-product initiates an interview-style interaction to define the product mission and roadmap.
• /shape-spec converts this plan into an MVP scope through targeted questions.
• /write-spec formalizes these discussions into a spec.md document.
• /create-tasks then breaks down the spec into a prioritized tasks.md list, grouped for implementation.

For code generation, Agent OS 2.1 offers two distinct approaches:

1. /implement-tasks: This method iterates sequentially through task groups, spinning up a new sub-agent for each, leading to faster execution.
2. /orchestrate-tasks: This offers greater architectural control by generating an orchestration.yml file, allowing specific agents to be assigned to particular task groups (e.g., front-end vs. back-end). For the iOS app, the "ios-developer" agent was assigned to all tasks.

Comparative testing revealed that the /implement-tasks approach was marginally faster due to less overhead, yet both methods yielded almost identical initial code quality. While basic CRUD functionalities were handled proficiently, complex UI elements like smooth scrolling and screen recording required approximately three rounds of manual refinement for each version to resolve initial glitches, demonstrating the current limitations for nuanced UI.

Pros of Agent OS:

• ✅ Context Efficiency: Sub-agents prevent context pollution, enhancing task-specific focus.
• ✅ Low Barrier to Entry: Well-documented, facilitating a quick learning curve for building new applications.
• ✅ Standard Management: Effectively automates the retrieval of coding standards via its "standards as skill" feature.

Cons of Agent OS:

• ❌ Bug Fixing Overhead: The full methodological process is cumbersome for minor bug fixes; direct prompting of Claude is often more efficient.
• ❌ Skill Injection Oversight: The system currently misses injecting the skill property into generated agent definition files, necessitating manual addition for sub-agents to utilize defined skills.
• ❌ Maintenance Velocity: With a single maintainer, the release cycle is comparatively slower than other rapidly evolving AI tools.

Compared to other frameworks, Agent OS is lighter than BMAD v4, generating more concise specifications, and its sub-agent functionality is currently more mature than BMAD v6 alpha. It offers iteration speed comparable to Spec Kit but is slower than Open Spec due to its emphasis on deliberate task creation. Uniquely, Agent OS automates the setup of Claude Code skills, a feature not intrinsically handled by other frameworks.

Final Takeaway: Ultimately, the underlying AI model (e.g., Claude Opus 4.5) profoundly impacts outcome more than the specific methodology. While frameworks like Agent OS serve as excellent pedagogical tools for understanding fundamental patterns such as agents, Multi-Context Processing (MCP), skills, and hooks within the Claude Code ecosystem, users should prioritize mastering these core platform tools. Obsessing over a particular framework's rules risks making the "process" the "product." Instead, the focus should remain on building the final application and leveraging foundational AI coding capabilities to adapt to any framework or directly prompt the model as needed.