The Harness Economy: Agents Build the Computer While Models Cool

Where it’s heading

The aggregate picture is an inversion. Active AI repos across sixteen categories gained roughly 747,000 stars this week, and about 529,000 of them — roughly 71 percent — landed in just two: Agents (303,265) and Coding Assistants (226,222). Concentration alone would make this an agent week. The accelerations make it a structural one.

Agents posted the largest positive acceleration of any sizable category (+15.14) on top of the largest absolute gain. The layers underneath are cooling at the same time: Language Models decelerated hardest (−15.17) even while recording the highest relative growth rate of any large category (0.00995/day) — the signature of a release spike fading rather than durable momentum. Chat Assistants (−11.64) and Inference · Serving (−5.08) are slowing too. Attention is climbing the stack. Weights and serving are becoming assumed; harnesses, memory, and tooling are where stars get spent.

The week’s feature coverage reads in the same direction. DeepSeek Harness sits at 242,904 stars, adding 9,707 a day, and the ecosystem forming around it has the texture of a platform economy: four thousand community plugins in seven weeks, an awesome-list that counts, sorts, and warns about them, and a Tauri desktop shell that replaces the npm ritual with a download. Around the agent, builders are assembling the rest of the machine — a Rust headless browser that renders only on demand (moli), virtual machines and drivers built for agent use (cua), Google’s distributed runtime that treats agent execution as a durable, resumable event stream (ax), a deployment skill that provisions real infrastructure with human approval at every write (golive-skill), and Cloudflare’s adversarial vulnerability harness. Memory is consolidating into its own layer, with two distinct bets on the table — write raw, spend at read time (nautilus-compass), and persistence-as-structured-reasoning (hindsight). Zhipu’s ZCode makes the thesis explicit, arguing GLM-5.3’s real value lives in the loop around the model; Anthropic’s claude-code demonstrates it, handling the bulk of its own maintenance. When infrastructure, memory, security, and deployment stories all orbit the agent, the direction isn’t ambiguous: open-source gravity has shifted from the model to the loop around it.

Categories on the move

Gaining: Agents (+15.14 acceleration, 303,265 stars), Other AI (+20.09, though on a 35-repo base), Image · Video · Audio (+6.36), LLMOps · Eval (+2.20), Learning (+2.17), and Creative · Design, which posted the highest relative growth rate of any category — 0.01676/day — on just 34 repos. Losing: Language Models (−15.17), Chat Assistants (−11.64), Inference · Serving (−5.08), Data Tooling (−1.62), ML Frameworks (−0.74).

So is it only coding agents? No — and the exceptions are the telling part. Coding Assistants held an enormous flat share, 226,222 stars at +0.32 acceleration: not speeding up, but not ceding anything. The categories rising behind Agents are its supply chain — eval tooling, learning resources, and image/video/audio, where VoiceStudio’s 2,226 stars a day is doing much of the category’s lifting. What’s decelerating is the commodity layer: model repos, chat wrappers, serving stacks, frameworks. That pattern suggests a maturing division of labor — the model is the given, the harness is the product, and eval is how you tell them apart. Creative · Design’s growth rate, small base and all, hints that generative tooling aimed at humans is carving a lane outside the agent boom.

Giants & breakouts

The heavyweight board has one center of gravity. DeepSeek Harness: 242,904 stars at 9,707 a day — more than four times the daily velocity of the number-two repo. VoiceStudio (52,619 stars, 2,225.6/day) is the only non-agent in the top four. anywhere-labs/dsh-desktop (29,886 stars, 1,904.6/day) pairs the highest relative velocity on the big board (0.06373/day) with the week’s clearest ecosystem signal — a double-click on-ramp to the harness. vectorize-io/hindsight (45,126, 1,836.7/day) is the memory bet with heavyweight traction; paperclipai/paperclip (96,763, 1,333.3/day) and browser-use/jev-ultrafast (21,701, 1,262.9/day, rel 0.05819) extend the agent run; stablyai/orca (84,455, 1,117.1/day) keeps Coding Assistants on the board; and rohitg00/ai-engineering-from-scratch (63,179, 1,056.4/day) shows Learning holding a top-eight slot on absolute velocity. Five of eight heavyweights are agents.

The newcomer board is stranger and more interesting. Edwardxlai/easyread leads on relative velocity (0.28220/day, 670 stars) out of Domain Apps, a category otherwise quiet. The middle of the board holds the week’s oddest sight: feder-cr/dots (2,574 stars, 625.1/day) and CopilotKit/OpenDots (2,540 stars, 582.3/day) — two repos with the same name, 34 stars apart, both agents, running neck and neck. KKKKhazix/AIHOT is the largest absolute newcomer (5,427 stars, 1,065.4/day) and lands in LLMOps · Eval, consistent with that category’s +2.20 acceleration. rehan-remade/universal-modder (2,641, 682.2/day) and nanaism/yomiyasu (1,320, 368/day) extend the coding-assistant run, and firelex/jeff (1,324, 245.2/day) deserves a pause: a Language Models newcomer breaking through in the category decelerating hardest — evidence that the model layer’s slowdown is saturation, not an absence of appetite for standout releases.

No prior-week boards are supplied, so the honest comparison is internal: nothing on either board is within reach of the harness, and nothing in the category data suggests that changes soon.