Agent workflows move from frameworks to marketplaces and an acquisition
DeepSeek opened a modular agent framework, GitHub brought outside agents into its marketplace, and SpaceX completed its Cursor purchase—three moves reshaping who controls software-building workflows.
Artificial Intelligence··Morning
DeepSeek opens Harness to developers
DeepSeek has published a developer preview of Harness, a framework that turns language models into agents able to drive external software, run code and finish complex tasks on their own. According to the South China Morning Post, its components are modular and swappable, so a programmer can reach directly into how the agent decides. Harness runs in four modes: a standard mode for general work, a code mode that executes commands across several applications, a creative mode that allows experimental custom tools, and a minimal mode for isolated testing. DeepSeek hired Cui Tianyu, previously at Jane Street, in March 2026 to lead its newly created Harness unit. The report writes that such tools are becoming an important layer in AI systems and are shifting the competition from model intelligence toward how seamlessly an agent connects to real-world software. In that framing, who can open, swap and audit the agent’s decision path matters as much as which model sits underneath. The preview hands that surface to developers rather than locking it inside a single closed product, presenting agent behaviour as a modular frame instead of a sealed service.[1]
GitHub pulls outside tools into the repository as agent apps
GitHub has opened agent apps, which pull outside tools into the repository workflow, through its Marketplace. The apps run on the same infrastructure as the company’s Copilot cloud agent. An agent app is called by assigning it to an issue, mentioning it in a comment under a pull request or picking it from the Agents tab. The first set includes Amplitude for product analytics, Endor Labs for dependency vulnerability scanning, LaunchDarkly for feature flags and PagerDuty for deployment risk assessment, alongside Packfiles, Miro, Bright Security, SonarQube and Octopus Deploy. GitHub’s stated aim is a single place where a developer can work without switching contexts. The move shows that the tools an agent can touch are no longer only local scripts; they can be chosen through an in-repo marketplace and call surface. Analytics, security scanning, feature flags and deployment risk can be triggered inside the same issue and pull-request thread, folding a vendor’s capability into the repository’s own flow. Control of the workflow therefore also rides on platform rules for which agent, with which tool set, may be invoked at which step.[2]
SpaceX closes Cursor at 60 billion dollars as workflow control redistributes
SpaceX’s purchase of the AI coding company Cursor took effect on 14 August, roughly two months after the agreement was announced. Business Today puts the price at 60 billion dollars. Cursor moves into the company’s SpaceXAI division. SpaceXAI combines Cursor’s software business with the company’s computing infrastructure and has so far shipped the coding-focused Grok 4.5 model, the task-running Grok Bot agents and Grok 4.6. At an employee meeting in August, Elon Musk said the company must succeed on the software front and that AI revenue would exceed all other SpaceX revenue by September. Cursor, founded in 2023, generates, edits and debugs code from prompts written in natural language. Read together with DeepSeek’s open framework preview and GitHub’s marketplace path for outside agent apps, the Cursor close shifts control of software-building workflows from model weights alone toward platform rules and product combinations. One move opens the agent’s decision layer to the programmer, another standardises third-party tool calls inside the repository, and the third folds a natural-language coding product into a large infrastructure and model family. Together they redistribute where an agent runs, which tools it may touch, and whose product line it stays inside.[3], [1], [2]
