obra/superpowers: The Agent Skills Ecosystem Explosion

On March 15, 2026, obra/superpowers is the single fastest-growing repository on GitHub: 84,570 stars with 1,893 added in the last 24 hours. It's a Shell-based "agentic skills framework and software development methodology." Not a model. Not a hosting platform. A way of composing agent capabilities.

That should make anyone building in the agent space pay attention. Here's why it matters, what it signals, and what it means for the projects we're building.

What Is Superpowers?

superpowers is deceptively simple: it's a framework for defining, composing, and executing agent skills using Shell scripts. The "methodology" part is the key differentiator—it's not just code, it's a structured approach to building agent capabilities that can be shared, versioned, and composed.

The core idea: instead of writing monolithic agent code or configuring complex YAML pipelines, you write small, focused Shell scripts that define discrete capabilities. Each skill is self-contained, testable, and composable.

The Trend Behind the Stars

superpowers isn't growing in isolation. The entire agent skills layer is exploding this week:

• agency-agents (45K stars, +31,758 this week) — a complete AI agency with specialized agents, each with "personality, processes, and proven deliverables"
• NousResearch hermes-agent (7.3K stars, +5K this week) — "The agent that grows with you," focused on personalization and learning
• deer-flow (bytedance, 30K stars, +5K this week) — SuperAgent harness with "sandboxes, memories, tools, skills and subagents"
• OpenViking (11.5K stars, +1.8K today) — context database specifically for agent memory and resource management

The pattern: developers are moving up the stack from "call an LLM API" to "compose agent capabilities from pre-built components." The question is no longer "can the model do this?" but "how do I wire together the right skills, tools, and memory?"

The Three-Layer Agent Stack

What we're seeing crystallize is a three-layer stack:

1. Model layer — LLMs (Claude, GPT, Gemini, open-source). Commoditizing fast.
2. Protocol layer — MCP, direct API calls, A2A. The transport mechanism for tools and data. Under active debate (see "MCP is dead; long live MCP" on Hacker News).
3. Skills layer — frameworks like superpowers, ClawHub, and DeerFlow that define what agents can do and how capabilities compose.

The skills layer is where the value is accumulating. Models are interchangeable. Protocols will standardize. But the library of well-tested, composable agent skills? That's a moat.

What This Means for promptengines.com

This trend directly intersects with our ecosystem. Here's the strategic read:

1. Skills are the new content

Just as blog posts and tutorials were the content that drove traffic in the SaaS era, well-crafted agent skills are becoming the content that drives adoption in the agent era. A skill that "just works" is worth more than a thousand-word tutorial.

2. The methodology matters more than the code

superpowers' explosive growth isn't because of its Shell scripts—it's because of the methodology for thinking about agent capabilities. The projects that define how to think about skills (not just what skills exist) will win.

3. Memory and context are the new frontier

With context windows plateauing at 1M tokens (Latent Space's "Context Drought" analysis from March 14), the real scaling path for agents is external context management. OpenViking's filesystem-based approach and cognee's "6 lines of code" pitch are both bets on this thesis. Any skills framework that integrates with persistent context systems will have an advantage.

The MCP Debate: Skills vs. Protocols

This week's Hacker News discussion ("MCP is dead; long live MCP," 175 points, 158 comments) is the other half of the story. The community is splitting into two camps:

Team MCP: Tools need deterministic, centrally maintained APIs with rapidly changing ground truth. MCP provides that. Usage is still booming despite the backlash.

Team Skills: Lightweight, local, natural-language procedures are sufficient for most use cases. MCP is over-engineered for what most agents need. Skills frameworks like superpowers prove this.

The pragmatic answer: both. MCP for enterprise integrations where you need type safety and contract enforcement. Skills for everything else. The frameworks that support both modes will win.

The Agent Memory Arms Race

The most technically interesting thread from this week's AI news: persistent memory and self-improvement. IBM's work on extracting reusable strategy/recovery/optimization tips from agent trajectories points toward systems that learn from their own execution history.

Three memory approaches competing:

• Vector databases (Chroma, Pinecone) — semantic search over past interactions. Mature but lossy.
• Context databases (OpenViking) — filesystem-based, hierarchical, self-evolving. Novel but unproven at scale.
• Structured memory (agent-internal files, JSON state) — deterministic, auditable, but manual. What most production agents actually use today.

superpowers and similar frameworks will need to pick a memory strategy. The ones that make it easy to plug in different backends will have an advantage over those that hard-code a single approach.

Practical Takeaways

1. Invest in the skills layer. The model and protocol layers will commoditize. Skills—well-tested, composable, documented capabilities—are where defensibility lives.
2. Start with Shell. superpowers' choice of Shell isn't accidental. It's the lowest-friction way to define agent capabilities. Consider whether your skills can be expressed as simple scripts before reaching for Python or TypeScript.
3. Plan for context overflow. 1M context windows are the ceiling for the foreseeable future. Your agent architecture needs external memory, not just longer contexts.
4. Watch the MCP stratification. Enterprise will use MCP. Everyone else will use skills. Build for both if you can, pick one if you must.
5. Methodology > code. The fastest-growing projects this week aren't the ones with the most features—they're the ones with the clearest approach. Define how to think about the problem, then ship the code.

Conclusion

obra/superpowers hitting 84K stars isn't about a Shell framework. It's a signal that the agent ecosystem is maturing past the "just wire up an API" phase into the "compose capabilities from a library of skills" phase. The projects that understand this shift—whether they're building skills, frameworks, or platforms—will capture the next wave of value.

The agent stack is consolidating. The skills layer is where the action is. And context management, not longer context windows, is the real scaling path.

---

Related Lab Notes:
The Context Drought: Anthropic Ships 1M in GA
Agent Memory Architectures
Agent Memory Is the New Context Window

References:
obra/superpowers: https://github.com/obra/superpowers
Latent Space AINews "Context Drought": https://www.latent.space/p/ainews-context-drought
Hacker News "MCP is dead; long live MCP": https://news.ycombinator.com/item (March 15, 2026)
GitHub Trending: https://github.com/trending