Drawgent is a single Rust binary that runs an Excalidraw editor and connects it to a user’s local coding agent (Claude Code, Codex or opencode) via the ACP bridge so the agent can read, edit and verify a live whiteboard. Setup verifies the agent CLI and login, installs ACP adapters into ~/.cache/drawgent/adapters (without bundled agent binaries), writes ~/.config/drawgent/config.toml, and ensures a headless Chrome renderer (offers a ~120 MB Chrome for Testing download or package-manager install). Claude and Codex bridges require Node.js ≥ 18 for their adapters. Typical workflow is drawgent setup <agent>, then drawgent up in a workspace (starts editor+API on 127.0.0.1:7300 or next free port and stores the scene in .drawgent/scene.json), or drawgent up attach to bind an existing session. Commands and options cover serve, mcp, room, port, token and permission policies.
Interaction includes a right-side chat panel and on-canvas AGENT: notes that fire after typing and carry position/context; the agent takes screenshots+scene, edits live, and resolves notes to DONE. Excalidraw.com rooms are supported as encrypted collaborators (AES-GCM, room Firestore storage) with the agent shown by a cursor. Agents expose MCP tools (get_scene, get_screenshot, add_elements, add_mermaid, update_elements, delete_elements, resolve_instruction, set_status) and an HTTP/WS API (/api/scene, /api/screenshot, /api/elements, /api/instructions, /api/chat, /ws). Build from source uses npm ci && npm run build and cargo install ., tests via cargo test and node smoke/e2e scripts. Known limits: renderer needs Chrome, Claude attach runs as a fork, Codex live attach is untested, one scene per workspace and images/files aren’t synced.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
OpenAI has paused training on its most advanced AI models due to ongoing concerns about unexpected and concerning behavior. The decision aims to address safety and ethical issues associated with these powerful models. (theverge.com)
DeepSeek Elastic Compute (DSec) is a system designed to optimize the allocation of elastic computing resources for deep learning workloads. It aims to improve efficiency and scalability in deploying large-scale AI models. (arxiv.org)
Spivak's Calculus has been formalized in Lean 4, covering every theorem and problem from the original text. The project aims to provide a comprehensive, machine-verified version of the classic calculus work. (github.com)
A potential burst of the artificial intelligence industry bubble could lead to significant financial and technological disruptions. Experts analyze the possible fallout and implications for the economy and innovation. (nytimes.com)
AI is often misunderstood as just math, but it involves complex processes similar to human cognition and has transformative societal impacts. The fear that AI will kill humanity within ten years is unfounded, as such scenarios are at least fifty years away and stem from human actions, not AI itself. (hughhowey.com)
Large language models (LLMs) are transforming software development, prompting questions about the source of truth in coding and how these models handle complex tasks at various levels of abstraction. Emphasizing the importance of understanding what is trusted as the foundation, the discussion explores the potential and limitations of LLMs in editing code across different programming layers. (intertwingly.net)
Today's best Hacker News stories, summarized and screenshotted, one email a day.