39K Stars for Agent Personas, 42K for a Skill Framework, 27K for a 1-bit LLM: What GitHub Trending in March 2026 Is Telling Us

During the second week of March 2026, something unusual played out on GitHub Trending. Three open-source AI agent projects from completely different categories surged to the top at the same time. One was a collection of AI agent personas, another was a methodology for directing coding agents, and the third was a framework that rethinks the inference efficiency of LLMs themselves.

The fact that three projects solving different problems exploded simultaneously was no coincidence. It was a signal: as AI agents move from experiments into real workflows, developers are simultaneously searching for answers to three questions — what to instruct the agent to do, how to instruct it, and where to run it.

1. agency-agents: Summon a Full Agent Team Instantly

msitarzewski/agency-agents¹ started with a single Reddit thread. Developer msitarzewski shared a set of AI agent personas, and within 12 hours more than 50 users were asking for a public release. The resulting repository now sits at 39,300 stars and 5,900 forks as of March 2026.²

The core idea is refreshingly simple: define each agent’s role as a markdown file, then drop it into your coding environment of choice — Claude Code, Cursor, Aider, Windsurf. Each agent file spells out identity, personality, core mission, task process, expected deliverables, and communication style.

The roster is modeled on a real agency. The engineering wing includes a frontend developer, a backend architect, a mobile app builder, and an AI engineer. Beyond engineering, you’ll find a UX researcher, a brand guardian, a growth hacker, a support responder, and a project shepherd. Standouts like “Whimsy Injector” and “Reality Checker” round out the more unusual corners of the lineup.

The bundled Nexus Spatial Discovery Exercise shows multi-agent collaboration in action. Eight agents — a product trend researcher, a backend architect, a brand guardian, a growth hacker, a support responder, a UX researcher, a project shepherd, and an XR interface architect — simultaneously evaluate a software opportunity and merge their outputs into a single report covering market validation, technical architecture, brand strategy, customer acquisition, UX research, and project execution.

Claude Code is the recommended default environment, but install scripts (convert.sh, install.sh) handle format conversion for other tools automatically. The MIT license means commercial use is fair game.

[!KEY] The real strength of agency-agents isn’t the framework — it’s the curation. Each agent file reflects months of community feedback and real-world refinement. Unlike a generic “act like an expert” prompt, these files define actual workflows alongside verifiable deliverables.

2. obra/superpowers: Spec-Driven Development Over Vibe Coding

obra/superpowers³ is a direct pushback against “vibe coding” — the practice of firing off loosely worded instructions at a coding agent and hoping for the best. It’s Jesse (obra)‘s systematized take on working with coding agents, refined over several months and released publicly in October 2025 when Claude Code launched its plugin system.⁴ By March 2026, the repo had cleared 42,000 stars,⁵ and in January 2026 it landed on Anthropic’s official Claude plugin marketplace.⁶

The philosophy is straightforward: don’t let the agent touch the code first. Instead, enforce this sequence:

1. Discovery: The agent figures out what you’re actually trying to build through conversation.
2. Spec writing: Specs are broken into chunks small enough that you’ll actually read them.
3. Review and sign-off: You approve the spec.
4. Implementation plan: A plan clear enough that “an enthusiastic junior engineer with poor judgment could still follow it.”
5. Sub-agent execution: Sub-agents handle individual engineering tasks, followed by a two-stage review — spec compliance check and code quality review.

YAGNI (You Aren’t Gonna Need It), DRY (Don’t Repeat Yourself), and TDD (Test-Driven Development) are baked in as defaults. The sub-agent system lets Claude work autonomously for hours without drifting off-plan.⁷

On the technical side, the skill system is the centerpiece. Agents load skill files on demand and act accordingly. Skills are composable, and the community keeps expanding the library. Supported platforms include the Claude Code official marketplace, Cursor plugin marketplace, Codex, OpenCode, and Gemini CLI.

Automatic Git worktree creation is another highlight. Each task gets its own isolated branch environment, so parallel workstreams don’t bleed into each other.

3. BitNet: Running a 100B-Parameter Model on a Laptop CPU

microsoft/BitNet⁸ operates on a different level entirely. It’s not about which agent to use or how to instruct it — it’s about running the model itself as efficiently as possible. First released in October 2024, the framework hit Hacker News’ front page again after a CPU inference optimization update on January 15, 2026.⁹¹⁰ Current star count: 27,400.

bitnet.cpp is an inference framework built specifically for 1-bit LLMs (technically 1.58-bit, using ternary values of -1, 0, and +1). Its kernels are optimized to run large models on standard CPUs — no GPU required. Official benchmarks:

The January 2026 optimization update added parallel kernel implementations and embedding quantization support, delivering an additional 1.15× – 2.1× speedup over the prior implementation.¹¹

The most striking number is the 100B scale. A BitNet b1.58 100B model runs at 5–7 tokens per second on a single CPU — roughly the pace of human reading. No GPU. No cloud. A 100B model running on an ordinary laptop.

Microsoft uploaded a 2B-parameter public model (BitNet-b1.58-2B-4T, trained on 4 trillion tokens) to Hugging Face in April 2025,¹² followed by a GPU inference kernel release in May 2025.¹³ An Azure-hosted demo¹⁴ lets you test it without building anything.

[!KEY] What BitNet proves is that “less bits, more capability” is achievable. 1-bit quantization isn’t just compression layered on top — it’s an architecture designed from training onwards. As this approach matures, practical LLM inference on edge devices without any cloud API becomes a realistic proposition.

Mapping All Three Projects on a Single Diagram

The three projects occupy distinct layers of the AI agent workflow.

graph TD
    User["Developer / User"]

    User -->|"What to instruct"| AA["agency-agents<br/>Agent Persona Library<br/>⭐ 39.3K"]
    User -->|"How to instruct"| SP["obra/superpowers<br/>Spec-Driven Development Methodology<br/>⭐ 42K+"]
    User -->|"Where to run"| BN["microsoft/BitNet<br/>1-bit CPU Inference Framework<br/>⭐ 27.4K"]

    AA -->|"Role definition"| Agent["AI Agent<br/>(Claude Code, Cursor, etc.)"]
    SP -->|"Execution methodology"| Agent
    BN -->|"Local LLM backend"| Model["1-bit LLM<br/>(On-device inference)"]

    Agent -->|"Model calls"| Model

agency-agents answers the what. Without a clear role definition, even a capable agent loses focus. A well-crafted persona file locks down the agent’s behavioral scope and expected outputs.

superpowers answers the how. A well-defined role still isn’t enough if the agent dives into code before the context is agreed upon. Spec-driven development forces alignment before implementation — enabling Claude to “work autonomously for a couple hours at a time without deviating from the plan.”

BitNet answers the where. Cloud APIs come with costs, latency, and privacy concerns. As 1-bit inference matures, the backend model powering an agent can move onto a local CPU.

The reason all three exploded at the same time is that these three questions aren’t independent. An agent only becomes genuinely deployable in real work when all three conditions are met simultaneously: a defined role, a sound methodology, and low enough execution cost.

Comparison: Which Project Fits Which Situation

The table below compares the use cases and entry barriers for all three projects.

What Remains After the Trending Rush

GitHub Trending is fickle. Plenty of projects rack up thousands of stars overnight and fade within a month. But the direction these three point to outlasts any trend cycle.

agency-agents is exploring a deeply human problem: giving AI agents a proper job description. Just as you’d define titles and responsibilities when staffing a team, there’s growing recognition that AI agents need the same treatment. This collection — born on Reddit and refined by community feedback over more than half a year — is the artifact of that process.

superpowers makes the case that good code from an agent requires good process. Its listing on Anthropic’s official marketplace signals that the company has formally endorsed this approach. If vibe coding is going to graduate from early experimentation into a real productivity tool, methodologies like this are non-negotiable.

BitNet has the longest time horizon of the three. The 1-bit LLM architecture is a software answer to hardware constraints. If a 100B model runs at 5–7 tok/s on a CPU, the economics of AI inference change fundamentally. An agent running locally — without a cloud API — offers a qualitatively different proposition in terms of privacy, cost, and latency.

For the broader picture of where the agent platform ecosystem is heading, see The Agent Platform War: NemoClaw, DeerFlow, and OpenClaw Compared. If you need a practical decision framework for choosing local LLM tooling, check out the Local LLM Tool Selection Guide.