Agent Framework Explosion — Superpowers, DeerFlow, NOMAD & TradingAgents

The week of March 20-26, 2026 saw an unusual concentration of new AI agent frameworks on GitHub. Four projects dominated the trending charts during that week. Each represents a different philosophy of what agents should do — from disciplined software development to deep research to offline survival to financial trading.

Note on numbers in this entry: weekly and total star counts throughout are as reported by 星探AI’s weekly “Agent 新战场” social-media posts at the time of each snapshot. They reflect a community ranking, not independently audited GitHub Star History data. For verification, look up each repo on Star History.

The Leaderboard

Stars gained (week of 3/20-3/26):

Superpowers  ████████████████████████████████████  +17,540
DeerFlow     ████████████████████████████████      +13,951
NOMAD        ███████████████████████████████       +13,848
TradingAgents ████████████████████               +8,939

#1 Superpowers — Teaching Agents Discipline

Most agent frameworks focus on what tools an agent can use. Superpowers focuses on how an agent should work — enforcing a structured software development methodology instead of letting agents freestyle code.

The 7-Stage Workflow

1. Brainstorm     → Refine ideas through questioning, not coding
2. Git Worktrees  → Create isolated workspaces
3. Write Plans    → Break into 2-5 min tasks with exact specs
4. Implement      → Subagent-driven or batch execution
5. TDD            → RED → GREEN → REFACTOR (enforced, not optional)
6. Code Review    → Assess against the plan
7. Branch Complete → Merge/PR options + cleanup

Why It Went Viral

The key insight: agents don’t fail because they can’t code — they fail because they skip the steps that make code work. Superpowers prevents agents from jumping straight to implementation, forcing brainstorming, planning, and test-driven development first.

Works with Claude Code, Cursor, Gemini CLI, and other coding agents. Ships with 14+ composable skills including systematic debugging (root-cause tracing, not guessing), parallel agent dispatching, and verification-before-completion.

#2 DeerFlow 2.0 — ByteDance’s SuperAgent Harness

DeerFlow (Deep Exploration and Efficient Research Flow) is ByteDance’s open-source SuperAgent framework — a ground-up rewrite from v1 that hit #1 on GitHub Trending the day it launched.

Architecture

┌─────────────────────────────────────────────────┐
│              DeerFlow 2.0 SuperAgent             │
│                                                  │
│  ┌──────────┐  ┌──────────┐  ┌──────────────┐  │
│  │ Planner  │  │ Sub-     │  │ Skills &     │  │
│  │ (task    │  │ agents   │  │ Tools        │  │
│  │  decomp) │  │ (delegated│  │ (extensible) │  │
│  └────┬─────┘  │  tasks)  │  └──────────────┘  │
│       │        └──────────┘                      │
│       ▼                                          │
│  ┌──────────┐  ┌──────────┐  ┌──────────────┐  │
│  │ Memory   │  │ Sandbox  │  │ Message      │  │
│  │ (long-   │  │ (Docker/ │  │ Gateway      │  │
│  │  term)   │  │  K8s)    │  │ (Slack/TG/   │  │
│  └──────────┘  └──────────┘  │  Feishu)     │  │
│                               └──────────────┘  │
│                                                  │
│  Models: Doubao-Seed-2.0, DeepSeek v3.2,        │
│          Kimi 2.5, OpenAI, Claude                │
└─────────────────────────────────────────────────┘

Key Features

Use Cases Beyond Research

The community has extended DeerFlow for data pipeline automation, presentation generation, dashboard creation, and content workflow management — far beyond its original research focus.

# Quick start
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow && make config    # Configure models
make docker-start              # Launch with Docker

#3 NOMAD — Offline Survival AI

Project N.O.M.A.D is the most unusual entry — a self-contained, offline survival computer. It bundles local AI, 99.6GB of Wikipedia, and critical tools into a Docker-based system that works with zero internet.

What’s Inside

Why It’s Trending

In a world increasingly dependent on cloud AI, NOMAD asks: what if the cloud isn’t there? It packages everything you’d need for AI-assisted decision-making into a system that runs on commodity hardware with no network. The intersection of prepper culture and AI captured developers’ imagination.

#4 TradingAgents — Wall Street in a Repo

TradingAgents simulates a professional trading firm using LLM-powered agents with specialized roles:

┌─────────────────────────────────────────┐
│         TradingAgents Framework          │
│                                          │
│  Analysts:                               │
│  ├── Fundamental Analyst (financials)    │
│  ├── Sentiment Analyst (news/social)     │
│  └── Technical Analyst (charts/signals)  │
│                                          │
│  Decision Layer:                         │
│  ├── Researcher (synthesize analysis)    │
│  ├── Traders (diverse risk profiles)     │
│  └── Risk Manager (position sizing)      │
│                                          │
│  Models: GPT-5.x, Gemini 3.x,           │
│          Claude 4.x, Grok 4.x, Ollama   │
└─────────────────────────────────────────┘

Supports multi-provider LLM backends (OpenAI, Google, Anthropic, xAI, OpenRouter, Ollama), backtesting with historical data, and a five-tier rating scale for trade recommendations. Treat it as a research and simulation framework, not financial, investment, or trading advice.

What This Wave Tells Us

1. Agents Are Specializing

The era of “general-purpose agent framework” is splitting into verticals: software development (Superpowers), research (DeerFlow), survival/offline (NOMAD), finance (TradingAgents). Each domain has unique requirements that generic frameworks can’t serve well.

2. Methodology > Tooling

Superpowers’ dominance (#1 with 130K+ stars) shows that developers care more about how agents work than what tools they have. The most impactful improvement isn’t adding more tools — it’s enforcing disciplined processes.

3. Open Source Is the Default

All four projects are open source. ByteDance (DeerFlow) open-sourcing their internal agent framework signals that even big companies see more value in community adoption than proprietary lock-in.

4. Offline AI Is a Real Category

NOMAD’s 13K+ star surge shows genuine demand for AI that works without internet — not just for preppers, but for edge deployment, privacy-sensitive use cases, and regions with unreliable connectivity.

May 2026 Update — Week of 5/02–5/08 Leaderboard (per 星探AI)

Six weeks later, the leaderboard had rotated. 星探AI published their agent-category top 5 for the week of May 2–8, 2026 (Episode #7 of their weekly “Agent 新战场” series). Numbers below are as reported by 星探AI’s social post and reflect their internal ranking method; treat as a community snapshot rather than independently audited GitHub Star History data.

Theme captured by 星探AI: “终端 + 金融 + 爬虫 + 编排” (Terminal + Finance + Crawler + Orchestration).

Note on ranking: 星探AI’s order isn’t pure weekly-gain — by that metric Skills (+~21K) would be first and TradingAgents (+~15K) second. Their published order seems to weight total + weekly together. For verifiable star history, check each repo on Star History.

What Each Project Is

TradingAgents — already profiled above. Still gaining stars six weeks later according to 星探AI’s snapshot. (repo)

Warp (github.com/warpdotdev/warp) — open-sourced its terminal client on April 30, 2026 under AGPL-3.0, with OpenAI listed as the founding sponsor. Marketed as an Agentic Development Environment (ADE) with native support for Claude Code, Codex, and Gemini CLI. The companion Oz platform orchestrates cloud agents from the terminal. Product site: warp.dev.

ruflo (github.com/ruvnet/ruflo) — previously named Claude Flow, renamed to ruflo by creator rUv. Multi-agent orchestration platform on top of Claude Code; ships with a large suite of MCP tools and pre-defined agent roles. The project’s own README claims notable SWE-bench numbers and API cost savings — useful as a marketing reference, not as an independent benchmark.

Scrapling (github.com/D4Vinci/Scrapling) — Karim Shoair’s adaptive Python web-scraping framework. Its agent-category interest comes from a built-in MCP server that lets Claude/Cursor agents extract structured content from web pages before LLM ingestion (cutting tokens and cost). The framework also advertises anti-detection features for sites with strong bot defenses — use responsibly and within each site’s terms of service.

Skills — mattpocock/skills (covered in its own wiki entry: Matt Pocock’s Skills — Claude Code for Real Engineers). Discipline skills for coding agents — TDD, structured debugging, architecture review. The largest weekly gain in the May snapshot.

What Changed Between March and May

Three signals worth noting:

1. Multi-agent orchestration is emerging as an infrastructure layer — ruflo’s traction suggests teams want a “swarm conductor” on top of Claude Code, not just better skills inside it.
2. Terminal is becoming a primary agent surface — Warp’s open-source pivot bets that the agentic future runs in the terminal as much as in the IDE. Aligns with Boris Cherny’s phone-based workflow direction.
3. Skills/discipline repos keep ranking high — six weeks apart, two different skill repos (Superpowers, mattpocock/skills) lead by weekly gains. The discipline-skills category isn’t a one-week story.

Links

• Superpowers: github.com/obra/superpowers
• DeerFlow: github.com/bytedance/deer-flow
• NOMAD: github.com/Crosstalk-Solutions/project-nomad
• TradingAgents: github.com/TauricResearch/TradingAgents
• DeerFlow 2.0 coverage: byteiota.com

How LearnAI Team Could Use This

• Use Superpowers-style workflows as a template for coding-agent lesson plans: brainstorm, plan, test, review, and finish branches deliberately.
• Use DeerFlow as a case study for long-horizon research agents with planning, memory, sandboxes, and messaging surfaces.
• Use NOMAD to discuss local-first AI, offline knowledge bases, and resilience for low-connectivity environments.
• Use TradingAgents as a cautionary finance-agent example focused on simulation, research, evaluation, and risk controls rather than investment advice.

Real-World Use Cases

• Software teams: Enforce repeatable agent development workflows with planning, TDD, review, and branch hygiene.
• Research and content teams: Coordinate long-running research, report generation, data analysis, and presentation drafts.
• Field and offline deployments: Run local AI plus offline reference libraries where internet access is unreliable or unavailable.
• Finance education and research: Simulate multi-agent market analysis for backtesting and decision-process study, with clear non-advice disclaimers.