Omarchy: The Linux Desktop Built for AI Agents

Omarchy 4.0.4 wires thirteen coding agents into Arch and Hyprland. Agent spend in the status bar, crashes routed to your agent, 367 self-documenting commands, and a shipped skill file that teaches any agent how the system is built.

Ruflo

Ruflo runs hierarchical and mesh agent swarms with Raft, Byzantine and Gossip consensus. MIT, 73,553 stars. It calls itself the original agent harness; OpenCode predates it.

claude-mem

claude-mem hooks the session lifecycle, compresses what your agent did and re-injects it next time. Works across many agents. Defaults to hosted memory, so set the local flag.

CodeGraph

CodeGraph parses your repository into a local SQLite knowledge graph and serves it over MCP to nine agents. 88% fewer tool calls, 62% fewer tokens, 100% local.

Herdr

Herdr owns the terminals your coding agents run in rather than wrapping them. Agent-aware pane state, survives a dropped SSH session, and a socket API agents drive themselves.

ZCode and the harness wars

Z.ai shipped ZCode, completing the set: every frontier lab now has a coding agent harness. All three lab-built ones have Issues disabled. None loads another s plugins.

DeepSeek Harness, 39 days later

In August we audited DeepSeek Harness and listed five things to watch. Five weeks and 66,000 stars later we re-ran the same queries. One prediction landed, and only halfway.

Qwen3.8-Flash-Next

Alibaba open-weight Qwen3.8-Flash-Next: a 125B MoE with 6B active parameters. What shipped, why the license is not Apache 2.0, the benchmarks, and the VRAM it actually needs.

DeepSeek Harness: The Open-Source Agent Layer Nobody Owns Yet

DeepSeek shipped the middle layer of the agent stack under MIT and told everyone to take it apart. Here is what dsh actually is, what the GitHub and npm APIs really say about its adoption, and why a free harness arrived three days before V4 output tokens got 4.5 times more expensive at peak.

DeepSeek-V4-Flash

DeepSeek-V4-Flash 0731 is official: MIT weights, 284B MoE (13B active), 1M context, and agent benchmarks above V4-Pro-Preview at $0.14/$0.28 per 1M tokens.

Mistral Medium 3.5

Mistral Medium 3.5 puts a 128B dense, multimodal, agentic-coding model on Hugging Face under a modified MIT license. Official 77.6% SWE-bench Verified, 256K context, and a $20M/month revenue carve-out that will not affect you. Here is the naming decoder, the honest benchmark picture, and the hardware it really takes.

LongCat-2.0

Meituan open-sourced LongCat-2.0: 1.6T parameters, about 48B active, native 1M context, MIT license. It spent two months anonymously topping OpenRouter agent leaderboards as Owl Alpha, and Meituan says it was trained entirely on domestic Chinese chips. Here are the numbers and their asterisks.

Qwen 3.6

Alibaba's Qwen 3.6 ships an open dense 27B that beats the old 397B flagship on coding and fits a 24GB GPU, under Apache 2.0. But the new flagship, Qwen3.6-Max, is closed. What you can download, and how to run it.

Kimi K2.7 Code

Moonshot's Kimi K2.7 Code keeps coding quality steady while thinking about 30 percent less, a cost story for long agent runs. The catch: the benchmarks are all first-party. What to believe, and how to run it.

OpenCode

OpenCode is the open-source coding agent for your terminal, model-agnostic, MIT-licensed, and the obvious landing spot after Google shut down Gemini CLI for consumer users. What it is, why it won, and how to point it at open models.