Meta paid $2B for an AI agent company whose secret weapon was... markdown files.
The Summary
- A developer reverse-engineered Manus's planning workflow — the AI agent company Meta acquired for $2 billion — into a Claude Code skill that uses persistent markdown for task orchestration
- The repo exploded to GitHub trending in under 24 hours and spawned at least 8 production forks, including finance agents, multi-project orchestrators, and entire Copilot agent frameworks
- The core insight: the competitive moat wasn't fancy proprietary tech, it was a file-based planning pattern anyone can copy
The Signal
Manus, the company Meta bought for $2 billion in January, built AI agents that could hold context across long-running projects. The industry assumed the value was in some proprietary LLM orchestration layer or custom training data. Turns out the real innovation was a workflow pattern: persistent markdown files that agents read and write to maintain state.
OthmanAdi's planning-with-files skill implements that exact pattern for Claude Code. When Claude starts a session, it reads a markdown plan. When it completes a task, it updates the plan. When the session ends, the plan persists. Next session, Claude picks up exactly where it left off. No vector databases. No complex state management. Just markdown files in a .codex directory.
"The competitive moat wasn't fancy proprietary tech, it was a file-based planning pattern anyone can copy."
The repo's viral trajectory tells you something about where agent development actually is versus where the hype cycle pretends it is:
- Trending on GitHub Python in under 24 hours
- 8+ production forks solving real problems (finance agents, multi-agent orchestration, GitHub repo audits)
- Multiple teams shipping full agent frameworks built directly on this pattern
- Version 2.34.1 already shipping cross-platform fixes and lifecycle hooks
What's notable isn't just that someone open-sourced a $2B workflow — it's that the workflow itself is radically simple. The skill uses three core files: PLAN.md for overall strategy, PROGRESS.md for session continuity, and CONTEXT.md for domain knowledge. That's it. No embeddings, no fine-tuning, no infrastructure moat.
The forks reveal what happens when you remove the complexity tax. @st01cs built an interview-first workflow with guaranteed command activation. @kmichels added multi-project support with git sync on session start. @Taoidle shipped multi-level task orchestration with parallel execution. These aren't incremental tweaks — they're entire agent architectures that wouldn't exist if the base pattern required a PhD to implement.
The economic signal here cuts deep: if a $2B acquisition's core value prop can be replicated in a weekend project and forked into production use cases within 24 hours, what does that say about moats in the agent economy? Either Manus had additional proprietary layers that aren't being discussed, or Meta paid $2B for execution talent and market position, not defensible technology.
The Implication
If you're building agents, strip out everything that isn't working. Manus won by making the state persistence problem boring — just markdown files in git. The complexity budget you save can go toward actually solving user problems instead of managing infrastructure.
For companies raising capital on "proprietary agent orchestration," this repo is a warning shot. The tools are commoditizing faster than pitchdecks can be written. Differentiation will come from domain expertise, distribution, and execution speed, not from architectural patterns that fit in 500 lines of code.
Watch what the forks ship next. That's where the real innovation happens — when developers can focus on workflow design instead of reinventing state management.