AI Agents as a Dev Team: Building a Mac Writing App in 1 Hour for $4.43
The author used mirasim to orchestrate an AI agent team (product, frontend, backend, QA, ops, code review) to build a Mac writing workbench (Electron/React/SQLite) in ~1 hour, consuming 82M tokens with 98% cache hit rate, costing $4.43 actual vs $19.15 estimated, delivering source code, 181 tests, 18 install checks, and independent code review.
The author describes using mirasim to organize an AI R&D team that built a macOS writing workbench called Wuming Draft in about one hour. The project was driven by the pain point of constantly switching between multiple coding agents (Codex, Claude Code, Qoder, Pi) and manually passing requirements, code, and review comments between them.
mirasim Workspace and Assistant Model
mirasim provides a shared workspace where multiple AI assistants can be configured with distinct roles, agents, models, skills, and MCP tools. Each assistant can be assigned a responsibility (e.g., product manager, frontend developer, code reviewer) and given access to specific local skills or existing agent subscriptions. A central Mira Bot routes tasks to the appropriate assistant or coordinates multiple assistants to work in parallel (e.g., having Claude Code and Codex produce two versions for comparison).
Team Composition and Collaboration Space
The author asked Mira Bot to assemble a full R&D team: R&D lead, product manager, frontend, backend, QA, ops, and an independent code review expert. After enabling the collaboration space, these assistants appeared automatically in the member list. The R&D lead was granted permission to invoke other assistants for task decomposition and progress tracking, while the code reviewer operated independently for security and quality checks. Different assistants could use different underlying agents (e.g., frontend switched to Claude Code) and draw quota from either mirasim's model service or the user's existing agent plans.
Project Requirements: Wuming Draft
The author provided a detailed prompt specifying a macOS desktop app for WeChat public account creators that organizes local Markdown files into a visual Kanban board. Key requirements included:
Tech stack: Electron, React, SQLite; read/write user's existing files; no custom cloud service or account system.
Four-state Kanban (To Write, Writing, To Publish, Published) with drag-and-drop, showing word count, image/layout status, update time, and stage dwell time; state change timeline; safe move/retract of published article directories; no overwrite on name conflict.
Conflict detection and backup on save/AI write-back; file operation path validation; default no upload of content, only transmit when user explicitly invokes AI or image hosting.
Editor: Markdown editing, live preview with scroll sync; paste images → upload via PicGo → insert link.
AI integration: local Codex and Qoder CLI for article Q&A, rewriting, generating "Moyu Green" public account HTML layout, one-click cover and inline image generation.
Role assignments: R&D lead decomposes and drives; PM defines requirements and acceptance criteria; frontend/backend implement and integrate; QA covers core flows and edge cases; ops handles packaging, install verification, rollback; code reviewer performs independent review.
Development Process and Human Intervention
After submitting the requirements to Mira Bot, the assistants took over task breakdown, implementation, handoff, bug fixing, and testing. The author could observe code changes, issues, and next assignees in the conversation. One manual intervention was required: the system requested screen recording permission (needed for local app verification), which the author granted before restarting the remaining tasks. Progress was trackable via an "Items" view with list, board, and table modes.
Deliverables and Verification
The final delivery included a runnable, installable local verification build with source code, test report, and usage instructions. The acceptance record shows 181 tests passed , 18 installation checks passed , and independent code review passed. Screenshots demonstrate the four-state Kanban populated with 250 articles from the author's actual public account directory, a three-pane writing workspace (file list, Markdown editor, live preview), and a standalone preview window for the "Moyu Green" layout with a "Copy to Public Account" button.
Cost Analysis: 82M Tokens, 98% Cache Hit, $4.43 Actual
The run consumed approximately 82 million tokens with a 98% cache hit rate . The usage page estimated $19.15 at official pricing, but the actual model cost was $4.43 — about 23% of the estimate. The author emphasizes that high cache hit rates dramatically reduce effective cost, and that token count alone is misleading; the real metric is actual quota deducted versus artifacts delivered.
Multiplayer Collaboration
After development, the author moved the project to a cloud collaboration space and invited colleagues. In the shared hall, multiple humans and multiple assistants collaborate simultaneously: humans can @ each other for discussion and @ different assistants for tasks. One contributor suggested making state transitions clickable on cards and enabling AI rewriting within the workbench, which were implemented. The author notes this blurs the line between human and AI team members, reducing handoff friction and making development more engaging.
Overall Experience
Using Codex or Claude Code directly means the human operates the tool; mirasim turns coding agents into team members, shifting human effort to clarifying goals and gating acceptance. Multiplayer mode puts humans and assistants in one room, keeping requirements, progress, and results in a single place, eliminating repetitive restatement. The system is not fully autonomous — OS-level permissions (like screen recording) still require human clicks. For developers who frequently switch between coding agents or want AI to follow a structured R&D process with role-based collaboration, mirasim is worth trying.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
