Open‑source agent stack

Software for artificial intelligence agents is being split into parts that developers can swap: one tool drives a browser, another runs code in a sandbox, another keeps models on-device. (github.com) (ai.google.dev) Vercel’s `agent-browser` is a command-line tool for browser automation with more than 50 commands for navigation, forms, screenshots, network and storage, plus isolated sessions for separate logins. Its GitHub repository showed about 28,100 stars on April 12, 2026, and the project says it works with coding assistants including Claude Code, Cursor, GitHub Copilot, OpenAI Codex and Gemini. (agent-browser.dev) (github.com) Google’s LiteRT-LM is the lightweight runtime in that stack: software that runs a language model directly on a phone, browser, desktop or Raspberry Pi instead of sending every request to a cloud server. Google describes it as open-source and production-ready for Android, iOS, web, desktop and Internet of Things devices, and its PyPI package reached version 0.10.1 on April 2, 2026. (ai.google.dev) (pypi.org) Other projects are focused on coordination. ByteDance’s DeerFlow says it orchestrates sub-agents, memory, tools, skills and sandboxes for tasks that can run from minutes to hours, and its repository said version 2.0 hit No. 1 on GitHub Trending on February 28, 2026. (deerflow.tech) (github.com) Rowboat is taking a different angle by treating an agent as a long-lived coworker with memory. The open-source desktop app says it connects to email and meeting notes, builds a knowledge graph from that work history, and acts on a user’s machine; its GitHub repository showed about 12,000 stars on April 12, 2026. (github.com) (rowboatlabs.com) Security has become a separate layer in the same stack because agents that write code often need permission to execute it. Nvidia said in a February 2026 security note that agentic workflows should block writes outside a workspace and restrict changes to configuration files, while LangChain introduced LangSmith Sandboxes in private preview on March 17, 2026, for running untrusted code. (developer.nvidia.com) (blog.langchain.com) A separate camp is trying to make email itself the control surface for agents. AgentMail says developers can create inboxes programmatically, send and receive messages, manage threads, and trigger workflows through webhooks and WebSockets, so an agent can operate through a normal email address instead of a custom dashboard. (agentmail.to) (docs.agentmail.to) That approach is aimed at a practical problem: many real-world tasks still arrive as messages, links, receipts, attachments and verification codes inside inboxes. AgentMail says its system exposes structured message data, attachments, labels and semantic search through an application programming interface, and Google’s Agent Development Kit lists an AgentMail integration for inbox creation, thread access and message sending. (docs.agentmail.to) (adk.dev) Investors are starting to treat that plumbing as a category of its own. TechCrunch reported on March 10, 2026, that AgentMail raised $6 million in seed funding led by General Catalyst, while Y Combinator lists the company as a Winter 2025 batch startup founded in 2025. (techcrunch.com) (ycombinator.com) The pattern across these projects is that “agent” is becoming less a single app than a bundle of interchangeable parts: browser control, memory, orchestration, model runtime and communications. Open-source developers are now publishing each layer separately, which makes the stack easier to inspect, replace and self-host. (github.com 1) (github.com 2)