GitHub Weekly: AI Agent Tools Trending Right Now (November 2 – November 8, 2026)

TL;DR: This week’s AI & Agents spotlight skews heavily toward the plumbing around coding agents rather than the agents themselves — plugin-based harnesses, spec-driven workflows, approval UIs, and infrastructure for agents that keep learning. The two biggest repos by stars are both pieces of DeepSeek’s open agent harness, DSH, and several others (Google, CopilotKit, a scrappy Mac menu-bar app) are all chasing the same question: how do you actually run, govern, and trust an AI agent once it’s doing real work for you.

Instead of the usual overall-trending list, this week we’re zooming in on AI and agent tooling specifically — repos tagged around autonomous agents, coding assistants, and the infrastructure that supports them. It’s a fast-moving corner of GitHub right now, and most of this week’s picks are brand new, having appeared only in the last three months. Here’s what’s getting attention between November 2 and November 8, 2026.

#1. deepseek-ai/deepseek-harness

github.com/deepseek-ai/deepseek-harness — TypeScript, 276,783 stars

DSH is DeepSeek’s own take on a coding-agent harness: the layer of software that sits between an AI model and your terminal, deciding how the agent reads files, runs commands, and talks to you. Its pitch is “everything is a plugin” — the editor integration, the model backend, even core behaviors are all swappable plugin modules rather than baked-in code, similar in spirit to tools like Claude Code or Cursor’s CLI but built to be taken apart and rebuilt.

Good for: developers who want full control over how their coding agent works under the hood, and plugin authors looking to extend or fork an agent harness instead of building one from scratch.

#2. Ebony-Vinyl/dsh-our-free-model

github.com/Ebony-Vinyl/dsh-our-free-model — JavaScript, 66,483 stars

This is a plugin for the DSH harness above that removes the usual friction of trying a new AI tool: no account, no API key, no usage cap. Install it and you get access to frontier-class models like DeepSeek V4.1 Flash and Kimi K3 for free, because the plugin handles the model connection for you behind the scenes.

Good for: anyone who wants to kick the tires on DSH without signing up for anything or paying for API credits first.

#3. google/artemis

github.com/google/artemis — Python, 17,051 stars

ARTEMIS takes plain-English instructions (“open the settings screen and toggle dark mode”) and turns them into real, repeatable Android automation — the kind of thing QA teams normally write brittle scripts for. It plugs into AI coding assistants like Claude Code, Codex, and Antigravity, and Google reports a 99%+ success rate on the AndroidWorld benchmark, a standard test suite for how well an agent can actually operate a phone.

Good for: QA and test-automation engineers who want to describe Android test flows in natural language instead of maintaining fragile UI-automation scripts by hand.

#4. google/elephant-goldfish-model

github.com/google/elephant-goldfish-model — Python, 15,496 stars

This one is an open-source plugin for “spec-driven development” — a workflow where you and an AI agent agree on a written specification before any code gets written, then use that spec to design and review the implementation together. It’s built to work across multiple coding agents (Claude Code, Codex, Gemini) rather than locking you into one vendor’s tooling.

Good for: teams that have been burned by agents going off and building the wrong thing, and want a structured checkpoint before code gets generated.

#5. lemomo-ai/lemo-opuscar

github.com/lemomo-ai/lemo-opuscar — JavaScript, 13,772 stars

A fun outlier on this list: a toolbox of 66 ready-made “film styles” for Claude Code, where each style is a prompt plus a demo short film generated entirely through code (think animated SVG or canvas sequences) rather than an actual video-generation model. It’s a showcase of using a coding agent as a creative-generation engine instead of just a code generator.

Good for: creative coders and anyone curious what a coding agent can do outside of writing application code — motion graphics, generative animation, and procedural art.

#6. CopilotKit/OpenBot

github.com/CopilotKit/OpenBot — TypeScript, 10,506 stars

OpenBot gives AI agents their own sandboxed “computer” — a browser, a filesystem, and a set of tools — and wraps every action in a governance layer that decides whether the action is allowed before it happens and logs what happened after. It’s built on AG-UI, an open protocol for connecting agents to user interfaces, so it isn’t locked to one specific agent framework.

Good for: teams building autonomous “AI coworkers” that need an audit trail and guardrails, not just raw capability.

#7. Human-Agent-Society/reef

github.com/Human-Agent-Society/reef — Python, 7,788 stars

Reef is infrastructure for agents that keep getting better from their own experience, using techniques like reinforcement learning and continual learning rather than relying purely on a fixed, pre-trained model. The goal is agents that improve the more they’re used, instead of staying static until the next model release.

Good for: ML engineers and researchers experimenting with agents that adapt over time in production rather than being frozen at training.

#8. XiaoDuoYa/codex-with-chatgpt

github.com/XiaoDuoYa/codex-with-chatgpt — TypeScript, 7,681 stars

This project lets you mix and match: keep OpenAI Codex’s tool-running harness (the part that actually edits files and executes commands) but swap in ChatGPT as the “planning brain” that decides what to do next, connected over MCP (Model Context Protocol, a standard way for models to talk to tools). It’s a neat example of decoupling an agent’s reasoning from its execution engine.

Good for: developers who like Codex’s execution loop but want a different model doing the actual thinking and planning.

#9. elie222/rakazo

github.com/elie222/rakazo — TypeScript, 7,600 stars

Rakazo is an open-source alternative to Grok’s built-in agent bot, but without the vendor lock-in — you choose which model powers it and which sandbox it runs its code in, instead of being tied to one company’s stack. It’s self-hostable, so you keep control of where it runs and what it can access.

Good for: teams that like the idea of a chat-driven coding agent but don’t want to be stuck with a single model provider.

#10. Louis-CFM/coucou

github.com/Louis-CFM/coucou — Swift, 7,068 stars

Coucou is a small Mac menu-bar app (living in the notch, if your Mac has one) and companion iPhone app that watches over your AI coding agents — Claude Code, Codex, Cursor, Gemini CLI, and more — and lets you approve their actions from the notch or your phone’s lock screen instead of having to keep a terminal window in focus.

Good for: developers running multiple coding agents at once who want a lightweight, always-visible way to catch and approve requests without babysitting a terminal.

Taken together, this week’s list says less about which model is smartest and more about the maturing ecosystem around agents — harnesses you can take apart, plugins that remove friction, governance layers that add trust, and even small conveniences like a notch app for approvals. Worth watching whether this “agent infrastructure” wave keeps picking up steam, or whether it consolidates around one or two dominant harnesses in the months ahead.