Open-source agent runtime that puts a research team in your terminal

What it does

FrontierAgent is a terminal-native framework for sustained research and file manipulation. It runs two primary workflows from the command line: a stateful ReAct agent that reads, writes, and executes inside a task-scoped sandbox, and an Agent Team mode where a coordinator delegates parallel sub-agents via a live task board. The same engine powers the bundled benchmark runner, so the workflows you use are the ones being measured.

The interesting bit

The architecture deliberately separates the generic loop, tools, workflows, and terminal UI into independent layers, meaning you can reuse the framework or benchmarks without dragging along the full TUI. The filesystem sandbox is fail-closed: /inputs is read-only, /workspace is scratch, and /outputs maps to a host directory alongside traces and checkpoints, making every run auditable and recoverable.

Key highlights

• Native Agent Team workflow with a live task board showing pending, active, completed, blocked, and cancelled states.
• Asynchronous intervention: queue new instructions while agents run; they inject at the next safe turn boundary without killing the active session.
• Sandboxed file and shell tools with approval gating for mutating operations, plus session checkpointing and /revert for recovery.
• Bundled evaluation harness with deterministic artifact collection and support for rerunning individual benchmark failures.
• Runs natively on macOS and Linux without Docker, or inside containers if preferred; model endpoint is pluggable via OpenAI-compatible APIs.

Caveats

• Local SGLang serving is tightly pinned to specific NVIDIA driver, CUDA, and SGLang versions; mismatches surface late as opaque CUDA or Triton kernel errors during model load.
• Windows support is WSL2-only, treated as Linux.
• Agent Team parallelism multiplies concurrent model calls, so benchmark concurrency settings need careful attention to avoid overwhelming endpoints.

Verdict

Worth a look if you want a reproducible, terminal-first agent environment for research or multi-step file tasks. Skip it if you need a polished web UI or are looking for a no-config desktop app.