ApodexAI/FrontierAgent

FrontierAgent – Open‑source runtime for autonomous AI agents

What it is – FrontierAgent is a Python‑based, terminal‑first framework that lets you run stateful LLM agents (ReAct style) or coordinated teams of sub‑agents on long‑horizon, file‑centric tasks. It ships a TUI (text UI) that shows a task board, sandboxed filesystem, live tool output, and a trace of every action.

Key capabilities

Feature	Why it matters
Two built‑in workflows – react (single agent) and agent_team (coordinator + parallel sub‑agents)	Choose a simple research loop or a decomposition‑and‑synthesis pipeline without writing code.
Sandboxed file system – /inputs (read‑only), /workspace (scratch), /outputs (deliverables)	Guarantees reproducibility and safety; all file operations are logged and can be reverted.
Asynchronous intervention – you can type new instructions while a run is executing; they are queued and injected at the next safe turn.	Enables interactive steering of long runs.
Approval & trace – mutating actions show a diff and require confirmation (or --yes). Full session checkpoints, logs, and a --resume flag let you pick up where you left off.	Prevents accidental data loss and aids debugging.
Evaluation harness – a subprocess runner that can execute dozens of research‑oriented benchmarks (BrowseComp, FrontierScience, GDPval, etc.) with deterministic artifact collection and parallelism controls.	Makes it easy to benchmark your own models or the bundled Apodex‑1.1 model.
Pluggable tools – web search (Serper or Parallel Search), shell, file manipulation, and custom “team” tools are provided out of the box.	Gives agents the ability to browse the web, run commands, and read/write files.
Model‑agnostic – works with any OpenAI‑compatible endpoint (Apodex‑1.1, hosted services, or self‑hosted SGLang).	No need to host the LLM yourself; just point OPENAI_BASE_URL to your endpoint.
Docker & native installs – pre‑built container images for Linux/ARM, optional GPU‑accelerated SGLang serving, or pure Python install via uv.	Flexible deployment on macOS, Linux, WSL2, or cloud GPU instances.

How you use it

1. Install – clone the repo and run uv sync (or uv tool install … for a global command).
2. Configure – put your API key and endpoint in .env (or a user‑wide config file).
3. Run – uv run frontier-agent --mode react --cwd /my/project for a single agent, or --mode agent_team to let the coordinator split work across parallel sub‑agents.
4. Interact – while the agent works, type follow‑up instructions; they are queued.
5. Inspect – the TUI sidebar shows the task board, live tool output, and a diff of any file changes.
6. Resume / revert – use --resume to continue a saved session or --revert to roll back file mutations.

Who it’s for – Researchers and developers who need a reproducible, controllable environment for autonomous agents that must read/write files, browse the web, or run shell commands. It is especially useful for evaluating new LLMs on complex, multi‑step benchmarks.

Documentation & resources –

• Full docs in docs/ (framework architecture, TUI guide, installation chooser, benchmark evaluation).
• Tech blog and arXiv report (Apodex‑1.1) linked at the top of the README.
• Community Discord, Hugging Face model collection, and a free two‑week trial of the Apodex API.

License – Apache 2.0.

---

All details above are taken directly from the repository’s README; no additional features are inferred.