GitAgent, Distributed AGI, and Research While You Sleep
·6 min read·ai-agentsdeveloper-toolsautonomous-systemssecuritymcp
For engineers, designers & product people. Stay up to date with free daily digest.
TLDR: GitAgent turns repos into portable agents, Hyperspace launches a distributed AGI swarm, and Claude Code gets an "auto-research in your sleep" upgrade.
If AI agents are the new microservices, today is basically the Dockerfile moment. Specs, skills, swarms: the stack is getting opinionated fast. Your job is deciding which opinions get to live in prod.
Key Signal
GitAgent: A git-native spec for portable AI agents
Hook: Your agent is now just three files and a git push away from a new framework.
What happened: The team behind GitAgent posted a Show HN describing an open spec that defines an AI agent as files in a Git repository. An agent is just three core files: agent.yaml (config), SOUL.md (instructions and personality), and SKILL.md (capabilities and tools). That definition can export to multiple ecosystems including Claude Code, OpenAI Agents SDK, CrewAI, Google ADK, LangChain, and others.
Why it matters: Agent projects are currently framework glue with a side of vendor lock-in. A portable, git-first spec lets you treat agents like code: versioned, reviewable, and migratable across runtimes. For you, this means you can design agents once, then target different execution platforms without ripping up your entire stack.
What to watch: Expect competing "agent spec" standards and conversion tools; you should pick something that keeps your core logic in plain text under git.
Hyperspace AGI: A peer-to-peer swarm for distributed "AGI" research
Hook: It is like SETI@home, except your CPU is gossiping with thousands of agents.
What happened: hyperspaceai/agi hit GitHub Trending with 500+ stars, pitching itself as "the first distributed AGI system". The project coordinates thousands of autonomous AI agents that collaboratively train models, share experiments via peer-to-peer gossip, and publish results back to the network. You can join from a browser or CLI and participate in the shared experiment pool.
Why it matters: Ignore the AGI branding and you still get something interesting: a decentralized orchestration layer where agents coordinate experiments and share learnings. It is closer to a distributed AutoML-plus-experiment tracker than a sci-fi brain. For you, this means you can start thinking about agent systems as networked research workers instead of single-host scripts.
What to watch: As of 2026-03-15, you should scrutinize data governance, reproducibility, and contribution quality before pointing this at anything with real IP or customer data.
ARIS: Claude Code that runs ML research while you sleep
Hook: Your new research intern never complains about hyperparameter sweeps.
What happened: wanshuiyin/Auto-claude-code-research-in-sleep, nicknamed ARIS, is trending with 1,000+ stars. It is a set of Claude Code skills for autonomous ML research: cross-model review loops, idea discovery, and experiment automation wired through Anthropic's Codex Model Context Protocol (MCP). The repo wraps workflows so Claude can propose experiments, run them, then critique results using multiple models.
Why it matters: This is a concrete template for turning a coding assistant into an actual research agent. Instead of manually iterating on experiments, you can hand off boilerplate design, execution, and first-pass analysis. For you, this means you can reserve human time for framing questions and vetting results rather than wiring yet another training loop.
What to watch: As of 2026-03-15, you should treat ARIS outputs as drafts, not ground truth; build guardrails around compute budgets, dataset access, and evaluation criteria.
Worth Reading 📚
A 400-page algorithms book finished with TypeScript and AI help
An engineer released a free ~400 page "Algorithms and Data Structures in TypeScript" via Show HN. The project started a decade ago in JavaScript and was recently revived, migrated to TypeScript, and completed using a spec-driven workflow with Zenflow and Claude Opus 4.6. The author provided structure and early chapters; AI generated most of the remaining content.
So what: You should treat this as both a solid TS reference and a living example of AI-assisted long-form technical writing.
734+ cybersecurity skills for AI agents, mapped to MITRE ATT&CK
The Anthropic-Cybersecurity-Skills repo packages more than 734 structured security skills for AI agents. Skills follow the agentskills.io standard and map to the MITRE ATT&CK framework, with ready wiring for Claude Code, GitHub Copilot, Codex CLI, Cursor, and Gemini CLI. It focuses on blue-team, cloud security, and digital forensics tasks.
So what: You should use this as a plug-and-play capability library when turning LLMs into actual security co-pilots instead of ad hoc chat prompts.
Signet: Autonomous wildfire tracking from open satellite and weather data
A Show HN introduced Signet, a Go-based autonomous system that handles the wildfire monitoring loop usually done manually. It ingests NASA FIRMS thermal detections, GOES-19 imagery, NWS forecasts, LANDFIRE fuel models, USGS elevation, Census population data, and OpenStreetMap. The agent filters detections, assesses conditions, and tracks fires worth monitoring.
So what: You should study this if you are building agents over messy geospatial or environmental data pipelines instead of clean SaaS APIs.
Model Context Protocol gets a mainstream explainer for the agentic era
The Next Web published "Rise of model context protocol in the agentic era". The piece focuses on MCP servers exposing tools and clients handling "elicitation" so agents can request parameters and maintain two-way interaction. Instead of hard-coded API calls, agents dynamically select tools and gather missing inputs.
So what: You should learn MCP (or an equivalent) if you want your agents to scale beyond single-shot RAG toys to real tool-using systems.
On the Radar 👀
Why Agentic AI is a Game-Changer for Ecommerce outlines how autonomous agents manage ecommerce workflows, fraud risks, and coupon abuse while keeping customer service reps in the loop.
The Rapid Trajectory Of Artificial Intelligence in Forbes highlights Klover.ai's "Artificial General Decision Making" multi-agent architecture that augments human decisions instead of replacing them.
GitHub’s slopocalypse and the cost of AI spam PRs shares Jannis Leidel's perspective, via Simon Willison, on how AI-generated pull request spam is breaking open membership models like Jazzband.
Agentic engineering fireside chat at Pragmatic Summit recaps Simon Willison's talk about phases of AI coding adoption, from copy-paste prompts to structured agentic systems.
New Tools & Repos 🧰
agi 578★: Fully peer-to-peer distributed AGI research network coordinating thousands of autonomous agents to train models and share experiments via gossip.
Auto-claude-code-research-in-sleep 1007★: Claude Code skills that automate ML research workflows with cross-model review loops and Codex MCP integrations.
Anthropic-Cybersecurity-Skills 617★: Library of 734+ structured cybersecurity skills for AI agents, mapped to MITRE ATT&CK and usable across major AI coding tools.
As of 2026-03-15 for all evolving technical claims and project statuses above.
Key Takeaways
- GitAgent defines agents as repos so you can move them across frameworks without rewrites
- Hyperspace AGI coordinates thousands of peer agents for distributed training and experiment sharing
- ARIS turns Claude Code into an autonomous ML research assistant with MCP-based tooling
- Standardized cybersecurity skills give you a plug-and-play library of blue-team agent capabilities
- Model Context Protocol is quietly becoming the glue layer for serious agent stacks
More from the Digest
For engineers, designers & product people. Stay up to date with free daily digest.