Munk AI Runs Open Source App Autonomously: 212 Requirements, 227k Lines, 24 Releases in 19 Days

Munk AI is a Harness Runtime for multi-agent collaboration (The harness runtime for multi-agent collaboration)

It organizes multiple "expert agents" into an AI-driven team that operates autonomously. You only make key decisions; Munk handles the product from start to finish and continues to operate it.

Over the past 19 days, to refine this system in a real scenario, I used the Public Build approach to let it take full control and autonomously iterate an open-source desktop application — OpenCompare (a free, open-source tool for comparing files and directories).

Currently, its core features are essentially complete.

A few simple numbers summarize the progress:

• 212 requirements closed: 213 requirements were received, 212 were successfully delivered and launched.
• 24 agent role assignments: Includes 6 core R&D decision roles (Product Owner, Planner for batch planning, Intake for requirement decomposition, Architect, Coding for development, Delivery for delivery review) plus 18 specialist acceptance and operations roles (GitHub sentry, CDP rendering acceptance, real-window acceptance, large-directory performance testing, daily release coordination, etc.).
• Only 20 human interactions; 50,861 autonomous agent interactions: Across 1,249 sessions and 50,881 total interactions, humans communicated with the Team Lead Agent (Product Owner) only 20 times (mainly for release approval, token strategy, and key directional decisions). The remaining 50,861 interactions (99.96%) were closed-loop internal agent-to-agent work: decomposition, coding, integration, cross-review, and real testing.
• 213.7 hours of autonomous work: 146.1 hours of development + 67.6 hours of operations and acceptance.
• 227k lines of code: 544 Git commits (519 agent-committed, 95.4% of the total), with 24 consecutive releases (v0.1.0 → v0.2.23). On GitHub's official contributor leaderboard, munk-agent stands as the top contributor.
• GitHub sentry on duty: Automatically replied to and followed up on 12 Issues (10 fixed and closed), automatically investigated and handled 24 CI errors, and produced 334 acceptance reports.
• 6.13 billion tokens: Cache hit rate 93.9% (DeepSeek consumed 6.018 billion for the primary development, Gemini 4 consumed 118 million for top-level decisions).

Latest interface screenshots

Real-world feedback and Munk AI 2.0 beta

During the iterative development of OpenCompare, we also exposed and fixed a number of problems in Munk AI itself, with corresponding optimizations:

• More stable: eliminated invalid retries in long chains.
• Token cost reduced by 50%: the current token consumption for the same work has dropped by half.
• Context strategy optimization: agents react more quickly, significantly reducing the hallucination rate caused by long contexts.

Based on this real-world iteration, Munk AI 2.0 is now in public beta. Feel free to download and provide feedback: 👉 Munk AI 2.0 download: https://github.com/munk-sh/munk-publish/releases

---

Open collaboration template (Open Company) and next steps

To keep dozens of agents working together consistently and stably over time, we distilled the goals, division of labor, and shift schedules of this multi-agent system into an open collaboration standard — Open Company.

The complete "open-source software maintenance team" template that we used to run OpenCompare is now open-sourced. Anyone can fork it or import it into Munk AI with one click: 👉 Multi-agent collaboration template library: https://github.com/munk-sh/open-company

Of course, Munk AI is not limited to maintaining open-source software; it can do much more—but every real scenario requires this kind of practical, dedicated polishing.

OpenCompare is Munk AI's first open-source product for desktop apps. Next, we'll continue crafting new team templates, exploring fields like mobile, XR, and self-media operations.

Which area would you most like to see Munk AI tackle next for team templates? Feel free to leave a comment or file an Issue!

We'll also start live-streaming soon on YouTube (24/7) and Bilibili (during daytime hours). Follow us in advance so you'll be notified when streams begin:

• 🔴 YouTube (24/7): youtube.com/@chriszhu5012/live
• 📺 Bilibili (daytime) : live.bilibili.com/31041209
• 📦 OpenCompare open-source repo: https://github.com/chaxiu/OpenCompare