Daily AI Research Briefing — March 15, 2026
Sunday edition. Anthropic's 1M context GA dominated the week, but GitHub's trending page tells a different story: the agent skills infrastructure layer is exploding. Here's what matters.
The Top 5
EXPLODING obra/superpowers: the skills framework eating agent dev
obra's superpowers is a Shell-based agentic skills framework and development methodology that's become the fastest-growing repo on GitHub today. At 84K+ stars with nearly 2K added in 24 hours, it represents a broader trend: developers want opinionated frameworks for agent skill composition, not just raw API access. The methodology angle matters—this isn't just code, it's a way of working.
DEBATE "MCP is dead; long live MCP" hits Hacker News
The MCP backlash isn't about demand—it's about ergonomics. Engineers are arguing that MCP tools are strong for deterministic, centrally maintained APIs but overkill for lightweight local procedures. Meanwhile, Chrome v146 ships native web MCP support, and LangChain demos a Deep Agent that continuously browses X via MCP. The protocol isn't dying; it's being stratified into "enterprise MCP" and "casual skills."
HOT OpenViking: context database purpose-built for AI agents
Volcengine's OpenViking unifies memory, resources, and skills management through a filesystem paradigm. It explicitly targets agents like OpenClaw and supports hierarchical context delivery with self-evolving capabilities. With context windows plateauing at 1M tokens (per this week's Latent Space analysis), external context databases like OpenViking are becoming the real scaling path for agent memory.
TREND Lightpanda: the headless browser built for AI
Lightpanda is a headless browser designed specifically for AI and automation workloads—not a stripped Chrome, but a purpose-built engine in Zig. At 1,323 stars today, it signals growing demand for lightweight, deterministic browser automation that doesn't carry the overhead of a full Chromium instance. For agent systems doing web scraping and browser-use tasks, this is a meaningful infra improvement.
RESEARCH Anthropic's 1M context GA — but the ceiling is real
Anthropic shipped 1M context in GA with strong MRCR benchmarks fighting "context rot." But Latent Space's swyx makes the contrarian case: context windows have been stuck at ~1M for two years, and physical HBM/DRAM constraints mean they won't meaningfully grow in the next 2 years. Prediction: "context rationing" is coming—free tiers may get 1K context windows while 1M becomes a premium tier.
📝 Today's Deep Dive
obra/superpowers and the agent skills ecosystem explosion
With 84K stars and the fastest growth on GitHub, superpowers represents a shift in how developers build agent capabilities. We're seeing the emergence of a skills layer between raw LLM APIs and end-user applications—a layer that promptengines.com sits directly on. Full article coming next.
GitHub Trending: The Agent Stack This Week
Looking at weekly trends, the pattern is clear: agents, agents, agents.
- agency-agents (msitarzewski) — 45K stars, +31,758 this week. A complete AI agency with specialized agents: frontend wizards, Reddit ninjas, reality checkers.
- Microsoft BitNet — 34K stars, +5,861 this week. Official inference framework for 1-bit LLMs. Efficiency is the new moat.
- deer-flow (bytedance) — 30K stars, +5,106 this week. Open-source SuperAgent harness with sandboxes, memories, tools, skills, and subagents.
- karpathy/nanochat — 48K stars, +3,628 this week. "The best ChatGPT that $100 can buy." Karpathy continues shipping.
- notebooklm-py (teng-lin) — 5.7K stars, +2,293 this week. Unofficial Python API for NotebookLM with full agentic skill support for Claude Code, Codex, and OpenClaw.
Latent Space Signal
From today's AINews: the agent infrastructure debate is crystallizing around three axes—memory (persistent, structured), tool access (MCP vs skills vs direct API), and orchestration (single agent vs multi-agent). The IBM work on extracting reusable strategy/recovery/optimization tips from agent trajectories points toward self-improving systems, while the MCP ergonomics backlash suggests the tool protocol layer needs simplification before it scales.
Also Noted
- Anthropic claude-plugins-official — Official directory of Claude Code plugins, showing Anthropic investing in its developer ecosystem
- NousResearch hermes-agent — "The agent that grows with you" — 7.3K stars, +5K this week. Personalization-focused agent platform
- cognee (topoteretes) — "Knowledge Engine for AI Agent Memory in 6 lines of code" — another memory solution entering the space
- InsForge — "Give agents everything they need to ship fullstack apps" — backend infrastructure purpose-built for agentic development
Sources:
GitHub Trending (daily + weekly), March 15, 2026
Latent Space AINews: "Context Drought" — March 14, 2026
Hacker News front page, March 15, 2026
Next briefing: March 16, 2026 at 1:00 PM UTC