Open‑source agents go mainstream

Open-source agent projects are shifting attention from picking a model to running software that can actually do work on a machine or in a browser. (github.com, ai.google.dev, github.com) An agent is software that can call tools, browse pages, read files, or run code instead of only generating text. Google says LiteRT-LM adds function calling for “agentic workflows” and runs on Android, iOS, web, desktop, and Raspberry Pi-class devices. (ai.google.dev) That execution layer now spans several distinct jobs. Rowboat says it connects to email and meeting notes to build a local knowledge graph, DeerFlow says it orchestrates sub-agents, memory, tools and sandboxes for tasks that can last minutes to hours, and Vercel’s agent-browser exposes browser automation through a command line interface. (github.com, github.com, agent-browser.dev) The recent burst of interest is visible on GitHub. As of April 12, 2026, DeerFlow had about 60,200 stars, agent-browser about 28,100, Rowboat about 12,000, and LiteRT-LM about 2,500. (github.com, github.com, github.com, github.com) The practical problem these tools target is reliability. A browser agent needs a stable way to identify buttons and forms, which is why agent-browser returns a compact accessibility tree with fixed references like “@e1” instead of relying on brittle screen scraping. (agent-browser.dev) The other problem is where the agent runs. Rowboat says it works “privately, on your machine,” while LiteRT-LM is positioned for offline, on-device deployment, including a Google showcase app that runs models locally on supported phones. (github.com, ai.google.dev) Security has become part of the stack rather than an afterthought. Agent Locker, an open-source sandbox project, says it runs agents inside isolated Docker containers with explicitly defined access to files, tools and commands instead of giving them direct access to the host machine. (github.com, devpost.com) LiteRT-LM shows how far that stack is moving below the chatbot layer. Google lists support for GPUs and neural processing units, multimodal models with vision and audio, and constrained decoding for tool use, with benchmark examples ranging from a Samsung Galaxy S26 Ultra to a Raspberry Pi 5 and a MacBook Pro with an M4 chip. (ai.google.dev) DeerFlow pushes in the opposite direction: longer, more composite jobs. Its repository describes a “SuperAgent harness,” and the project said version 2 reached the top spot on GitHub Trending on February 28, 2026, after adding more skills, tracing options and Kubernetes sandbox provisioning. (github.com) The pattern across these projects is that the model is only one part of the system. The rest of the work now sits in runtimes, browser control, memory, and containment layers that decide whether an agent can finish a task without breaking the machine it runs on. (github.com, github.com, github.com, github.com)