frontieragent

FrontierAgent is an open-source agent framework and terminal product for long-horizon research and file-based work. It provides a native command-line TUI with two workflows: a stateful ReAct agent for focused research, repository analysis, and document work, and an Agent Team mode where a coordinator maintains a task board, delegates parallel sub-agents, collects reports, and synthesizes results. The system includes a task-scoped sandbox with read-only /inputs, working /workspace, and persistent /outputs directories, plus approval gates, local tracing, checkpointing, session resume, and /revert recovery. It supports asynchronous intervention, allowing users to queue instructions during active runs. The repository also includes a benchmark evaluation harness supporting datasets such as BrowseComp, Humanity's Last Exam, GDPval, and FrontierSearchBench. Installation uses Git, Python 3.12, and uv, with optional Docker and SGLang support for local models on NVIDIA GPUs. The project is licensed under Apache 2.0.