Daily Briefing
15 August 2026 · 5 stories
🔥 Top story
1DeepSeek ships a build-it-yourself kit for AI agents
Imagine a digital Lego set where every piece (the brain, the tools, the memory, the test zone, the interface) snaps on and off. Chinese lab DeepSeek just released such a kit, called Harness, for free under the MIT licence, which means anyone can use and modify it. In two days the project racked up nearly 99,000 stars on GitHub, a sign developers worldwide are paying attention. Practically, it’s a direct alternative to Anthropic’s Claude Code, but designed the other way round: you start with the modules, and the tool communication standard adapts afterwards. Four modes are ready to use, including a minimal one that DeepSeek itself uses to test its own models. For a team wanting to assemble an agent without reinventing the wheel, this is a fresh starting point.
Source : https://www.deepseek.com/harness/en/
2Alibaba releases a compact, free version of its best agent model
After the disappointment of the giant, paid version on 12 August, Alibaba ships one that fits on a Mac: Qwen3.8-27B, 27 billion dense parameters, Apache 2.0 licence (commercially usable, free of charge). It reads images and text, holds context over 262,000 words (up to one million as an option), and scores 84.3% on a benchmark where it must operate a computer like a human. That beats the previous generation (73.3) and beats Anthropic’s Opus 4.6 Max (72.7). For a small business or freelancer who wants a powerful model without sending data to a giant cloud, this is a real option you can use today.
Source : https://huggingface.co/Qwen/Qwen3.8-27B
3Writer unveils an AI agent team 52% cheaper and that watches its own spending
Writer overhauled the engine of its agents: a new model version (Palmyra X6) and a new orchestrator (the harness). The result: 52% cheaper and 48% faster on the same tasks. The model itself is a specialised fine-tune of a Chinese model (Z.ai’s GLM 5.2), but entirely retrained on American data centres for legal reasons. On the practical side: 2 dollars per million words read, 8 per million words written, and agents hold up for eight hours without supervision. The real novelty: administrators finally see token consumption as a budget line, not a surprise. A meaningful step in the price war that actually hides rising consumption.
Source : https://writer.com/blog/palmyra-x6-major-harness-release/
4An open standard so AI agents keep memory without leaking your secrets
Picture a company filing cabinet where any agent can come read and drop notes, but nothing ever leaves the company. That’s what Starling Memory Works proposes with its Universal Cognitive Architecture (UCA), published for free under Creative Commons, no fees. The AI model itself stays memoryless: it forgets everything once the task is done, like an employee who returns their files when leaving. All notes stay in plain text in a space controlled by the organisation. This is the fourth major agent-memory building block in four weeks, after AWS, LangChain and Tencent: how the agent remembers is becoming as strategic as which model it uses.
Source : https://www.starlingmx.com/free-standard
5GitLens 19 turns the code history into a shared kitchen for humans and agents
GitKraken releases GitLens 19, which turns the traditional Git commit history into a workspace where humans and agents collaborate. Three modes: Review (inspect before proposing a change), Compose (reorganise into clean commits), Repository-wide context (see at a glance what touches a file). Compatible with VS Code, Cursor and Antigravity. The 2026 report released at the same time reveals one striking number: 28% of developers now say their main way of working is “autonomous AI”, up from 7.6% nine months ago. The market no longer sells isolated agents; it sells the environment where several agents live with humans.
Source : https://www.prnewswire.com/news-releases/gitkraken-introduces-gitlens-19-bringing-human-and-ai-workflows-together-in-one-workbench-302851472.html
📡 To watch
Agent memory and engine layers mature in parallel
Four memory blocks (AWS, LangChain, Tencent, Starling) in four weeks, and four open-source agent engines in seven days (DeepSeek Harness, Writer X6, Microsoft Agent Framework, NVIDIA NOOA): both families are maturing at the same speed. A break in one ripples through the other. Worth watching to anticipate the next move.
US-side retraining of Chinese models is becoming a stated strategy
Writer openly admits it picks up GLM 5.2 (Z.ai) from Hugging Face, retrains it on US data centres and publishes a technical report. This is the written response to the 12 August criticisms about Z.ai lacking a safety framework. Any company that wants to use a Chinese model without falling under Chinese jurisdiction now takes this path. A reproducible pattern, not a one-off.
Qwen3.8-27B’s scores depend on the engine used to measure them
The official card specifies that the code tests are run with Claude Code, Anthropic’s engine. That’s like rating a car with the competitor’s engine under the bonnet: the number stays good, but it reflects the combination. Cross-check with an independent benchmark before any production-critical use.
Two visions of the agent clash on the same day
On 13 August, DeepSeek (Chinese, open-source) and Microsoft (American, managed) each published their agent framework. One problem, two opposite answers: open code versus managed product, community-first versus enterprise-first. The match is on. For users, a philosophical choice is coming in the next months.
📊 Trend
This fortnight, the invisible layers of agents (memory, engine, distribution) are maturing as fast as the models themselves. Chinese labs now ship a complete stack (model + framework + memory + distribution), not just weights to download. On the American side, US-side retraining of Chinese models becomes a stated practice to stay compatible with regulations. The playing field shifts: the battle is no longer over the model’s raw performance but over the whole ecosystem each vendor can deliver.