Nova — The Agent's Computer

Nova is a full-stack computer agent that researches, codes, and creates. Instead of a plain chat window, it gives you the "Agent's Computer" — a live 6-tab panel (Terminal, Editor with live diff, Browser preview, Activity timeline, Files, Review) streaming exactly what the agent is doing in real time. Built by Ali Technologies on top of the open-source DeerFlow harness, Nova adds +35,738 lines across 338 files verified against the upstream tag. Highlights: a verify loop where the agent tests its own builds via headless-Chromium checks (console errors, blank renders, screenshots) with a vision path so the model sees its output; a deterministic, no-LLM code review engine giving plain-English verdicts plus developer risk flags; self-correction middleware with iteration budgets and dead-end detection; 32 agent tools; and an 11-probe self-healing ops watchdog. For this hackathon, Nova serves inference on AMD Instinct GPUs as a first-class, one-click choice. Settings → Models offers two AMD paths: Fireworks AI (managed, on AMD Instinct MI300X) and AMD Developer Cloud, running Nova's own vLLM-on-ROCm inference. Switching is configuration, not code. Usage is independently verifiable via GET /api/models/amd-usage, with a conservative AMD badge shown only for verified AMD-hosted endpoints. Track 1 adds a lean, token-efficient Fireworks batch harness agent, buildable and smoke-testable with one command. Nova also ships a full skills/tools system, a sub-agent architecture for parallel task decomposition, persistent long-term memory, IM channel support (Telegram, Slack, Feishu, WeChat, WeCom, DingTalk) with no public IP required, and 8,112 lines of new tests across 37 files.