DeepSeek open-sources Harness — everything is a plugin

Share
DeepSeek open-sources Harness — everything is a plugin

DeepSeek is moving up the stack from models to the software developers actually use, and Uber is scaling robotaxis across Europe. Two open-source and mobility moves to start your Friday.

DeepSeek released DeepSeek Harness (dsh), an open-source agent harness under the MIT license in developer preview, built on an architecture where "every capability is a plugin" — from the model adapter and tool registry to the session log and the agent loop itself. It's the lab's first serious product beyond the model layer: a React-based web app that VentureBeat sizes up as an open-source rival to Claude Code, capable of spawning Claude Code and Codex agents through their SDKs and shipping default modes like code-mode with programmatic tool calling and bash+edit. The harness is powered by the Cordis meta-framework, whose design is spelled out in a paper on composable programming paradigms, and the plugin model means teams can swap any component without forking the codebase. Version 0.1.0 is explicitly rough — DeepSeek's Tianyi Cui warned of "many rough edges" and promised compatibility-breaking changes — yet the repo had already crossed 85,000 GitHub stars within a day of going live.

Why it matters: this is DeepSeek executing the open-source playbook a second time in 48 hours, hot on the heels of DeepSeek's V4 Pro open weights finally land on Hugging Face. But the bigger signal is strategic — the company is expanding beyond models into the harness layer where agent developers actually get locked in. The "everything is a plugin" bet is the interesting part: if it sticks, DeepSeek isn't just undercutting frontier labs on price, it's offering an open alternative to the closed agent stacks of OpenAI and Anthropic, with the plugin ecosystem as the moat.


Uber and Pony.ai expanded their robotaxi partnership, planning to deploy more than 2,000 Pony.ai robotaxis across Europe — adding four cities beyond the existing Zagreb service, with a Middle East expansion to follow. Under the "joint-deployment model," Pony.ai supplies its L4 self-driving technology and operational know-how, Uber provides the platform — booking, payment, customer service — and local partners handle day-to-day fleet operations, with vehicle ownership able to sit with different partners in each market. The companies haven't named the cities or a timeline, saying details will be revealed in phases. Pony.ai already runs paid, fully driverless services in China's four tier-one cities with city-wide breakeven economics in multiple markets, and the pair launched Europe's first commercial robotaxi service in Zagreb with Croatian firm Verne earlier this year. It's the latest milestone in a fast-moving week for Pony.ai, which also saw Pony.ai's fourth-gen driverless heavy truck enters mass production.

The move sharpens the industry's two competing strategies: Uber's aggregation play, now spanning more than 30 autonomous vehicle partners, versus the vertically integrated fleets of Waymo and Tesla. "The next chapter for autonomous mobility is about moving from individual launches to repeatable commercial scale," said Uber's Sarfraz Maredia — and Europe's fragmented regulation makes the multi-partner model the pragmatic path there, even if it cedes control of the vehicle itself.

What to watch: whether Harness's plugin bet attracts an ecosystem of its own — and which European cities Uber names first.

DeepSeek is betting the harness, not the model, decides where developers get locked in — will "everything is a plugin" win that fight? Write to us via our contact page.

Sources: DeepSeek Harness (GitHub)

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Cursor launches Origin, a GitHub rival for agents

Cursor launches Origin, a GitHub rival for agents

The AI coding tool that just sold for sixty billion dollars is done borrowing someone else's git host. Cursor began rolling out Origin today — its own code-hosting service — in early beta on every paid plan, with repos, pull requests, code browsing, and live GitHub sync in the first cut. Agent-native features, the company says, ship soon. Origin lives in a new Codebase tab inside the editor. Create a repo, push, and the code is hosted on Cursor.

DFlash 2 brings 20% faster parallel drafting to the inference stack

DFlash 2 brings 20% faster parallel drafting to the inference stack

Inference is eating AI's bill, and a new open-source release just made it cheaper to run. DFlash 2, a block-diffusion draft model from Inco AI, ships checkpoints for Qwen3.8-27B and Muse-Glimmer-30B — and claims over 20% more output per verification pass at roughly 1% added latency. The core idea is speculative decoding: a small draft model guesses a block of tokens, the large model verifies the whole block in one forward pass.

Mojo compiler goes open source under Apache 2.0

Mojo compiler goes open source under Apache 2.0

Modular has thrown open the source for Mojo, the systems programming language it built for the AI-accelerator era, releasing the full compiler and toolchain under the Apache 2.0 license (with LLVM exceptions) on GitHub. For a language that spent four years in public development behind a closed compiler, this is the unlock that turns Mojo from a vendor experiment into something the wider community can actually own. Modular has open-sourced the entire Mojo compiler and toolchain under Apache…

Minitap says Google's Artemis shipped its code with the credit stripped

Minitap says Google's Artemis shipped its code with the credit stripped

Google's Pixel test team open-sourced an Android automation agent this week. The team it was built on top of says the authorship trail was rewritten before anyone saw it. Minitap says Google's Artemis contains its code, and that the names of the people who wrote it were removed. Minitap published mobile-use in August 2025 under Apache 2.0 — an agent that drives real Android and iOS apps, 2,821 GitHub stars, and the reference implementation behind a clutch of mobile-agent papers.

Hugging Face runs async RL training with no NCCL and no weight ship

Hugging Face runs async RL training with no NCCL and no weight ship

The hardest part of reinforcement learning is not the algorithm. It is getting a trainer and a bunch of inference servers to keep passing weights back and forth without one of them stalling. Hugging Face just published a recipe that does it with three rented jobs, one storage bucket, an HTTP proxy, and no NCCL at all. Hugging Face trained a reasoning model end to end using async GRPO and LoRA, and the whole weight sync is a file upload.