Open‑source agent ecosystem grows

Open-source agent projects are clustering into a simple stack: run a model locally, give it tools, then fence it in. (youtube.com) An agent is software that can plan steps and use tools, not just answer a prompt. In the latest roundup tied to the YouTube video published as episode No. 247, the projects named were Rowboat, LiteRT-LM, DeerFlow, agent-browser and Locker. (youtube.com) Those projects line up around different jobs. Google’s LiteRT-LM is a C++ runtime for running language models on edge devices, and Google’s documentation shows it serving models locally for desktop agents through a separate `lit` server process. (ai.google.dev, adk.dev) Rowboat handles memory and workflow context. Its GitHub page says the software connects to email and meeting notes, builds a long-lived knowledge graph, and runs privately on a user’s machine; the repository showed about 12,000 stars on April 12, 2026. (github.com) DeerFlow handles orchestration, which is the layer that breaks a task into smaller jobs and routes them to tools or sub-agents. ByteDance’s repository describes DeerFlow as an open-source “super agent” harness with memory, skills, sandboxes and sub-agents, and the project page said version 2 reached No. 1 on GitHub Trending on February 28, 2026. (github.com, deerflow.tech) agent-browser covers browser control, which is the part that lets an agent click, type and navigate websites. Vercel Labs describes it as a browser automation command-line tool for agents, and its site says it returns compact accessibility-tree snapshots instead of raw pages to save tokens. (github.com, agent-browser.dev) Locker addresses containment, the security layer that keeps an agent from roaming across a full machine. The Agent Locker repository says it runs agents inside isolated Docker containers with model and tool access defined inside the sandbox instead of on the host system. (github.com) The pattern is visible in the numbers. On April 12, 2026, DeerFlow’s public repository showed about 60,200 stars, agent-browser about 28,100, Rowboat about 12,000, and LiteRT-LM about 2,500, suggesting builders are putting attention into separate layers rather than one monolithic agent app. (github.com, github.com, github.com, github.com) Google’s own examples show why the runtime layer is getting attention. Its LiteRT-LM overview lists offline demos on Android, iPhone, Linux, macOS and Raspberry Pi 5, with the same system exposing Python, Kotlin and C++ interfaces for developers. (ai.google.dev) That leaves open-source agents looking less like one product and more like Lego bricks. The projects in this week’s roundup point to a stack where one tool runs the model, another remembers context, another drives the browser, and another keeps the whole thing inside a box. (youtube.com, github.com, github.com, github.com, github.com, github.com)