DeepSeek open-sourced a plugin harness. Meta shipped a 30B model that fits on a laptop GPU. Docker walled off YOLO mode. Claude Code sessions started talking. OpenAI put Codex on Linux.
The desktop agent stack stopped being a thesis this week. It started shipping parts.
DeepSeek dropped an open harness. Meta shipped a 30-billion-parameter model that fits on a consumer GPU. Docker put a microVM around the coding agents people already run with the safety switch off. Anthropic let Claude Code sessions talk to each other on the same machine. OpenAI finally put Codex on a Linux desktop. None of that moves the model frontier. All of it is infrastructure for the category this publication spent the last seven days arguing into existence.
DeepSeek Harness landed Thursday as a developer preview, MIT-licensed, source included, and it picked up 78,612 GitHub stars before Friday morning. The repo did not exist Wednesday. The pitch is the argument August has been making, written as a product: the model is not the agent. The harness is. DeepSeek’s version of that idea is more aggressive than most. Every capability is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI all mount through a kernel called Cordis. You swap any of them in configuration without touching the harness source.
Treat it as a design document you can run, not a runtime you should bet a production desk on. DeepSeek says so in the README. Compatibility-breaking changes are coming. The part worth stealing today is the bookkeeping. Every run writes an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, every context injection. Resume, fork, search, and replay all operate on that stream. We spent Thursday walking through Claude Code’s append-oriented session storage as a reference implementation. A major lab shipped the same pattern as open source, with a Trajectory view that lets you inspect the record by source.
Four modes come with it. Standard is a full coding agent. Code mode lets the model write TypeScript that orchestrates multiple tool rounds in one program. Minimal keeps a shell and a file editor so you can bench the model without the rest of the furniture. Creator mode is the one I would actually install: inspect the current runtime, test Cordis plugins in memory, and combine them into new modes. That last piece is how you tell this is a harness for people who build harnesses, not a chatbot with extras bolted on.
A local runtime still needs a model that can live on the machine. Meta Superintelligence Labs released Muse Glimmer for that job: 30 billion parameters, Apache 2.0, open weights on Hugging Face as meta-models/Muse-Glimmer-30B. The card is dated August 9. By midweek it had more than 120,000 downloads. Full precision would take over 55 GB. The K-Quant-17GB release compresses the weights to roughly 4-bit, under 20 GB, which leaves room for the KV cache, a perception encoder, and a speculative-decoding drafter inside a 24 GB or 32 GB envelope. That is a MacBook M4 Max or an RTX 5090, not a cluster.
They measured the K-Quant-17GB build with a DFlash drafter. Speculative decoding made decode 3.1 times faster on an RTX 5090, 1.8 times on an M5 Max, and 1.5 times on an M4 Max. Meta compared it to Gemma 4 31B and Qwen 3.6 27B and claimed strong results for the size class on agentic work, including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. I am not going to recite a vendor table as if I ran the evals. The claim that matters for this week is architectural: a lab that spent two years pushing people toward the cloud shipped a model designed to stay on your desk, trained for long-horizon execution, tool calling, and instruction following. Goose, Hermes, BeeAI, and OpenClaw all got a local brain that fits the hardware they already assume.
The personal agent stack we laid out Wednesday had an execution sandbox as a required layer. Docker spent the week making that layer a product. Docker Sandboxes isolate coding agents in a dedicated microVM with your project workspace mounted in and the host left alone. The agent can install packages, change configs, and start its own Docker containers. The machine you type on does not absorb any of that. The pitch is blunt: YOLO mode, safely. --dangerously-skip-permissions is how people actually run Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro when they want the agent to finish the job. Docker’s argument is that the answer is not more permission prompts. The answer is a wall.
I have mixed feelings about the marketing and none about the topology. Approval fatigue is a measured failure. Anthropic’s own auto-mode analysis found users approve about 93 percent of permission prompts. We covered that yesterday in the Claude Code harness writeup. A sandbox that assumes the human will click yes is the correct security model. The honest limit is the same one every isolation product hits: the workspace you mount is still the workspace you care about. A microVM that can rewrite your repo is safer than a process that can rewrite your home directory. It is not a review.
Anthropic shipped a smaller piece I have been waiting for since Tuesday’s category post. Claude Code v2.1.224 and later, on macOS and Linux, can list and message your other Claude Code sessions on the same machine. A message is a piece of text. It is not conversation history and it is not files. When one session lands a schema change that breaks another, Claude can warn the other session before you notice. Local delivery goes over a per-session socket. It never goes through Anthropic’s servers. Cross-machine and web sessions go through Anthropic. Permission boundaries stay per-session. Claude is instructed not to ask another session to do something its own rules would block.
Tuesday I wrote that nobody had solved the handoff off your laptop. This is not that. Your phone still cannot resume the agent. What it does solve is the other failure mode of the desktop category: five terminals, five agents, and you as the integration bus. I have been that bus. It is a terrible architecture. A named session you can mention with @ (v2.1.232) is a primitive, not a multi-agent framework. It is the right primitive.
OpenAI put Codex in the ChatGPT desktop app for Linux, now in preview. Ubuntu 24.04 and 26.04 LTS, Debian 13, Fedora 43 and 44, x64 and ARM64, .deb and .rpm. The app is a workspace for projects, files, browser workflows, and Codex sitting next to ChatGPT. A reply in the community thread attributed feature parity with Windows to Tibo. Early reports are the usual preview tax: Wayland IME needs extra flags, Arch and Nix users are staring at .deb packages, and Codex CLI projects on the same machine do not show up in the desktop project list. That last bug is the interesting one. The vendor desktop still cannot see the local runtime you were already using.
Goose, BeeAI, Hermes, and OpenClaw spent this week proving the open-source side of the category. OpenAI shipping a Linux desktop does not change the architecture. It confirms the demand. When the closed vendors start treating your machine as the agent’s environment, the argument that agents belong on the desktop is no longer a take. It is the market.
The pattern across the five is the stack assembling in public, one layer at a time. DeepSeek open-sourced the harness kernel. Meta shipped a local model that can sit inside it. Docker productized the sandbox. Anthropic connected the sessions. OpenAI showed up on Linux. That is the personal agent stack from Wednesday, arriving as five unrelated releases instead of one product. The category still does not have a single thing you can install. It has fewer missing pieces than it did on Saturday.
Next week the arc shifts from the desktop to multi-agent coordination: supervisor, hierarchical, swarm, blackboard, and the honest line between a decomposable workflow and token theater. Cross-session messaging is the hinge. Once sessions can talk, people will try to turn that into a team. Some of those teams will be real. Most of them will be expensive.
If this was useful, forward it to one engineer who needs less noise in their feed.