Four updates for people who build with AI landed on 8 and 9 October 2026. A software-delivery company bought the coding agents of a well-known AI coding start-up, but not the start-up itself. Hugging Face released a training library update that it says fits much longer texts on one graphics chip. Google declared its Dart framework for AI apps stable. And NVIDIA showed an AI tool for reviewing Linux kernel patches.
Harness buys select Augment Code assets
Confirmed (company releases); this is an asset purchase. Harness, which sells tools for testing and shipping software, said on 8 October that it “has acquired select Augment Code assets, including Cosmos and the Auggie CLI products, the Code Context Engine, and related technology”. The team behind those products “will join Harness” [1]. Harness’s headline says “Harness Acquires Augment Code”, but the text, and Augment’s own post, describe a sale of selected assets. Augment’s Matt McClernan wrote: “We’ve agreed to sell these select assets” [2].
No price was given. The New Stack reports the deal was “for an undisclosed amount” (reported) [3].
Cosmos becomes the “Harness Cosmos Software Factory Agent”. Harness says a “fleet” of Cosmos agents can plan a change, write the code and tests and open a pull request, with “each agent” working “in its own isolated VM”, and an engineer brought in for decisions such as approving a design or the final merge [1]. Teams can adapt prebuilt “Experts” called Project Builder, PR Author, Deep Reviewer and PR Fixer [1].
Harness chief executive Jyoti Bansal said: “AI’s impact on software will not be measured by how much code it generates, but by how much valuable software reaches customers” [1].
Not shipped yet (plan). Harness plans to connect Augment’s Code Context Engine to its Software Delivery Knowledge Graph, so agents can draw on past test failures, security findings and incidents. The New Stack reports those capabilities “have not shipped”, with no date given, and that Cosmos remains reachable through Augment’s existing site for now [1][3].
Hugging Face TRL v1.15.0: longer training sequences
Confirmed release; the figures are the maintainers’ own. TRL is Hugging Face’s open-source library for fine-tuning language models. Version 1.15.0, published on 8 October, makes a “fused LM head” the default in its main trainers. It avoids building one very large table of numbers during training. The headline claim: “up to 6.9× longer sequences on the same GPU”, and it is “on by default” [4].
The maintainers’ own tests ran on one NVIDIA B300 capped at 79 GiB of memory, with a small Gemma model (gemma-3-1b). The longest trainable sequence rose from 10,240 to 59,392 tokens for DPO (5.80×) and from 9,216 to 63,488 for KTO (6.89×). At 8,192 tokens, “peak memory drops 52% to 82% and steps are 2.3% to 10.9% faster” [4]. The notes say the gain “is largest for small models with large vocabularies” [4].
Breaking changes. A PEFT adapter on the model’s lm_head “now raises” an error, and use_liger_kernel=True is deprecated in four trainers [4].
Google’s Genkit Dart reaches 1.0
Confirmed. Google called Genkit Dart 1.0 “the first stable, production-ready release of Google’s open-source framework for building AI-powered features and agents in Dart”, the language behind Flutter apps [5]. It works with Gemini, Anthropic’s Claude, OpenAI and OpenAI-compatible models through one interface, and adds human-in-the-loop “interrupts” (pausing a tool, such as a hotel booking, until a person confirms), middleware, Dotprompt prompt files, and OpenTelemetry tracing via genkit_otel [5].
Experimental. Stateful agents and A2UI generative interfaces sit under a separate “experimental” import [5].
Also noted: NVIDIA’s Boro
Reported. Phoronix reports that NVIDIA’s Andrea Righi presented Boro, an AI-assisted tool for reviewing and testing Linux kernel patches, at the Linux Plumbers Conference in Prague [6]. The Apache-2.0 Rust repository has existed since June; this was a presentation, not a new release [6][7].
What this does not prove
- That Harness bought Augment Code the company. The releases describe a sale of “select” assets [1][2].
- What Harness paid. The price was not disclosed [3].
- That Cosmos already uses Harness’s delivery data. That link is planned and has not shipped [3].
- That every TRL user gets 6.9× longer sequences. These are the maintainers’ benchmarks on one GPU and one small model. The notes say they did not measure H100 hardware, flash-attention, real generation or multi-GPU set-ups [4].
- That Genkit Dart’s agent features are stable. Stateful agents and A2UI remain experimental [5].
The Bottom Line
Harness now owns Augment Code’s Cosmos, Auggie CLI and Context Engine, plus the team, at an undisclosed price, and the most interesting integration is still to come. TRL v1.15.0 promises much longer training runs on the same hardware, by its maintainers’ measure. Genkit Dart is stable at 1.0, with agents still experimental.
Related on TSN: AI Coding Tools 2026: Cursor, Windsurf, Copilot Compared; Same Model, Different Score: Training AI Inside the Coding Tools People Actually Use
Sources
- Harness, “Harness Acquires Augment Code, Advancing the Autonomous SDLC from Idea to Production”, press release, San Francisco, 8 October 2026 (company release). https://www.harness.io/press-and-news/harness-acquires-augment-code
- Matt McClernan, “Augment Code’s next chapter”, Augment Code blog, 8 October 2026 (company post). https://www.augmentcode.com/blog/augment-code-next-chapter
- Amanda Caswell, “Harness bought Augment’s coding agents. The best feature hasn’t shipped yet.”, The New Stack, 8 October 2026 (trade press; source of “undisclosed amount” and “not shipped”). https://thenewstack.io/harness-augment-cosmos-acquisition/
- Hugging Face, TRL v1.15.0 release notes, GitHub, published 8 October 2026, 20:26 BST (maintainers’ benchmark). https://github.com/huggingface/trl/releases/tag/v1.15.0
- Chris Gill, “Announcing Genkit Dart 1.0: Build production-ready agentic apps with Dart and Flutter”, The Flutter Blog, 8 October 2026. https://flutter.dev/blog/announcing-genkit-dart-1-0
- Michael Larabel, “Boro: NVIDIA’s Open-Source Effort For AI-Assisted Linux Kernel Development”, Phoronix, 9 October 2026 (trade press). https://www.phoronix.com/news/NVIDIA-Boro-Linux-Kernel-AI
- NVIDIA, boro repository, GitHub (Apache-2.0, Rust; created 23 June 2026; checked 9 October 2026). https://github.com/NVIDIA/boro
, Hugging Face's TRL v1.15.0 claims up to 6.9× longer training sequences, Google's Genkit Dart reaches 1.0, and NVIDIA shows its Boro kernel-review tool. What each says, and what it doesn't.){kind=link}