{"slug": "apodex-frontieragent", "title": "Apodex – FrontierAgent", "summary": "Apodex released FrontierAgent, an open-source agent runtime, terminal product, and evaluation suite for long-horizon research and file-based work, available through an OpenAI-compatible Apodex-1.1 endpoint with no model hosting required. The `frontier-agent` TUI ships two native workflows, ReAct and Agent Team, and the same workflow engine powers the benchmark runner used to evaluate Apodex models. The framework, tools, workflows, and evaluation layer remain separate so each can be reused independently.", "body_md": "FrontierAgent is an open-source agent runtime, terminal product, and evaluation\nsuite for long-horizon research and file-based work. The `frontier-agent` TUI\nships two native workflows:\n\n- **ReAct** — one stateful agent researches, reads files, writes deliverables,\nruns commands, and iterates in a task-scoped sandbox.\n- **Agent Team** — a coordinator maintains a task board, delegates independent\nwork to parallel sub-agents, collects their reports, and synthesizes the result.\n\nThe same workflow engine powers the benchmark runner used to evaluate Apodex models. The framework, tools, workflows, and evaluation layer remain separate, so each can be reused independently.\n\nImportant\n\n**No model hosting required.** Get an API key, connect to the OpenAI-compatible\nApodex-1.1 endpoint, and start running FrontierAgent in minutes.\n\n[→ Start building for free at platform.apodex.ai](https://platform.apodex.ai/)\n\n⏳ This is a limited-time offer—come try it and let us know what you build!\n\nNew here? Use the **[documentation index](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/README.md)** to find the right\ninstallation, SGLang, workflow, evaluation, or developer guide.\n\n- **Native Agent Team workflow.** The coordinator decomposes the request,\ndispatches bounded parallel assignments, receives structured reports, and can\nuse an optional fast reporter for final evidence review.\n- **Task Board.** Agent Team's`add_task` and`update_task` events appear live in\nthe TUI sidebar with pending, active, completed, blocked, and cancelled state.\n- **Sandboxed file work.** Shell and file tools share one task-scoped filesystem:`/inputs` is read-only,`/workspace` is working state, and`/outputs` contains\npersistent deliverables. Authorization and sandbox failures are fail-closed.\n- **Asynchronous intervention.** Type while an agent is running to queue a new\ninstruction. It is injected at the next safe turn boundary without discarding\nthe active run. In Agent Team mode it steers the coordinator; already-running\nsub-agents are allowed to finish.\n- **Transparent deliverables.** On macOS/Docker,`/outputs` maps to`.apodex/runs/<session-id>/outputs` on the host. The same run directory also\ncontains its checkpoint, trace, engine log, and trajectories.\n- **Approval, trace, and recovery.** Mutating operations show a diff and require\napproval unless`--yes` is enabled. Sessions are checkpointed, every action is\ntraced locally,`/revert` restores session changes, and`--resume` continues a\nsaved run.\n- **Evaluation included.** The subprocess runner supports research and\nfile-grounded benchmarks, deterministic artifact collection, concurrency,\nprogress inspection, and rerunning individual failures.\n\n*Conceptual Agent Team workflow, from task delegation and asynchronous report collection to verification and final synthesis.*\n\n``` php\nflowchart LR\n    U[\"User / benchmark task\"] --> TUI[\"TUI or subprocess runner\"]\n    TUI --> R[\"Stateful ReAct\"]\n    TUI --> C[\"Agent Team coordinator\"]\n    C --> B[\"Task board\"]\n    B --> S1[\"Sub-agent 1\"]\n    B --> S2[\"Sub-agent 2\"]\n    B --> SN[\"Sub-agent N\"]\n    R --> FS[\"Task sandbox\"]\n    S1 --> FS\n    S2 --> FS\n    SN --> FS\n    FS --> I[\"/inputs (read-only)\"]\n    FS --> W[\"/workspace (working files)\"]\n    FS --> O[\"/outputs (deliverables)\"]\n    S1 --> C\n    S2 --> C\n    SN --> C\n    C --> A[\"Final answer / report\"]\n    R --> A\n```\n\nThe repository boundaries are intentional:\n\n```\nfrontier_agent/  generic loop, scheduling, registries, AgentBus, observers\nplugins/tools/   web, shell, file, sandbox, and team tool implementations\nworkflows/       ReAct and Agent Team pipelines, profiles, prompts, observers\napodex/          terminal CLI/TUI, approvals, sessions, traces, and Docker path\nbenchmarks/      public harness plus bundled FrontierSearchBench/FrontierChallenge\n```\n\nMore detail: [framework architecture](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/framework.md),\n[Agent Team](https://github.com/ApodexAI/FrontierAgent/blob/main/workflows/agent_team/README.md), and\n[Stateful ReAct](https://github.com/ApodexAI/FrontierAgent/blob/main/workflows/stateful_react_agent/README.md). See\n[run artifacts and timestamps](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/run-artifacts.md) for the on-disk layout.\n\nRequirements: Git, Python 3.12, [uv](https://docs.astral.sh/uv/), and an\nOpenAI-compatible model endpoint. Docker is optional.\n\n```\ngit clone https://github.com/ApodexAI/FrontierAgent.git\ncd FrontierAgent\n\nuv sync --python 3.12 --extra dev\ncp .env.example .env\n```\n\nAdd your endpoint to `.env`:\n\n```\nOPENAI_API_KEY=your-key\nOPENAI_BASE_URL=https://your-openai-compatible-endpoint/v1\nOPENAI_MODEL=your-model-name\n\n# Optional web research tools\nSERPER_API_KEY=\nJINA_API_KEY=\n```\n\nStart the TUI:\n\n```\n# Stateful single-agent workflow\nuv run frontier-agent --mode react --cwd /path/to/project\n\n# Coordinator plus parallel sub-agents\nuv run frontier-agent --mode agent_team --cwd /path/to/project\n```\n\n`uv sync` above installs the lightweight terminal runtime. Scientific and\ndocument packages are intentionally optional in native mode; the agent installs\nonly what a task actually needs into `<project>/.apodex/runtime/native`. The\n`apodex` command is retained as a compatibility alias.\n\nPrefer a script that does all of the above? `./scripts/run-macos.sh` and\n`./scripts/run-linux.sh` set up a hosted-endpoint install, and\n`./scripts/run-linux-gpu.sh --install-system-deps --setup-only` prepares a native,\nisolated SGLang environment on a Linux NVIDIA GPU. The step-by-step equivalent is\nthe [endpoint quickstart](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/tui-endpoint-quickstart.md)\n([中文教程](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/tui-endpoint-quickstart.zh-CN.md)), which requires neither\nmodel self-hosting nor Docker.\n\nLocal SGLang serving is pinned to reviewed NVIDIA driver / CUDA / SGLang tracks,\nand a mismatch surfaces late as opaque CUDA or Triton kernel errors during model\nload. Confirm your `nvidia-smi` driver against the\n[GPU compatibility matrix](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/gpu-compatibility.md) before choosing an\nimage tag or native pin. The GPU helper selects a reviewed userspace track from\nthe host driver, but never installs or replaces the driver itself.\n\nThe operating system, FrontierAgent runtime, and model runtime are independent\nchoices. “NVIDIA” describes the local model service, not how the agent itself\nruns. Unsure which applies to your machine or GPU provider? Start with the\n**[installation chooser](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/README.md)**.\n\n| Environment | FrontierAgent runtime | Model endpoint | Start here | \n|---|---|---|---|\n| macOS | native or Docker Desktop | hosted or another OpenAI-compatible endpoint | [macOS](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/macos.md) | \n| Linux host/VM | native (default), bubblewrap, or Docker | hosted, native SGLang, or Docker SGLang | [Linux](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/linux.md) | \n| managed Linux GPU container | native inside the provider container | custom GPU image or native SGLang | [GPU platforms](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/gpu-platforms.md) | \n| Windows | WSL2, treated as Linux | hosted or a WSL2-reachable endpoint | [Linux/WSL2](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/linux.md#windows-and-wsl2) | \n\nChinese-speaking macOS users can use the\n[macOS 中文安装与一键启动指南](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/macos.zh-CN.md).\n\nPre-built `linux/amd64` and `linux/arm64` images are published to the GitHub\nContainer Registry, so no local Python environment is needed:\n\n```\ncp .env.example .env\ndocker compose run --rm agent\n```\n\n- [Run FrontierAgent in Docker](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/docker.md) — Compose, image\npinning, direct`docker run` , and EC2/ECS deployment.\n- [Docker SGLang on a Linux NVIDIA host](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/linux-nvidia.md) — two\ncontainers on one network; SGLang owns the GPU.\n- [Native SGLang without nested Docker](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/linux-nvidia-native.md) —\nfor managed GPU environments that forbid a nested daemon.\n- [SGLang configuration reference](https://github.com/ApodexAI/FrontierAgent/blob/main/config/sglang/README.md) — every`.env.sglang` variable, token-budget invariants, and tuning order. Production 35B templates\nfor RTX 4090, RTX 5090, and two-GPU hosts live under`config/sglang/` .\n\nRun without a task for an interactive session, or pass one and stay in the session for follow-ups:\n\n```\nuv run frontier-agent --mode agent_team --cwd /repo \\\n  \"Research the alternatives, verify the evidence, and write a report\"\n\n# One-shot, line mode, or resume a saved session\nuv run frontier-agent --mode react --cwd /repo -p \"explain src/main.py\"\nuv run frontier-agent --mode agent_team --no-tui \"compare these implementations\"\nuv run frontier-agent --resume\n\n# Attach read-only documents before the TUI starts (repeatable)\nuv run frontier-agent --mode react --cwd /repo \\\n  --input ~/Downloads/claim.pdf --input ~/Desktop/photo.jpg\n```\n\nThe sidebar carries the plan/task board, live tool activity, deliverables, and a session-scoped diff. While a workflow is busy, typing a follow-up queues it for the next safe turn boundary rather than interrupting the run.\n\nFor the four sidebar tabs, previews, approvals, attachments, clipboard support,\nkeys, and Agent Team live steering, see the\n[TUI user guide](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/tui-user-guide.md)\n([中文使用教程](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/tui-user-guide.zh-CN.md)). The full slash-command, option,\nand theming reference is [`apodex/README.md`](https://github.com/ApodexAI/FrontierAgent/blob/main/apodex/README.md).\n\n| Mode | Best for | Execution model | \n|---|---|---|\n| `react` | focused research, repository analysis, document/file work | one stateful agent using the `tui` workflow profile | \n| `agent_team` | broad questions that benefit from decomposition and parallel investigation | coordinator, persistent task board, bounded parallel sub-agents, report collection, synthesis | \n\nAgent Team parallelism is additional to benchmark concurrency. When evaluating,\nstart with `--concurrency 1`; total simultaneous model calls can approach runner\nconcurrency multiplied by the team spawn limit.\n\nSet `SWARM_NO_WEB=1` to disable Agent Team web tools or `REACT_NO_WEB=1` for\nclosed-book ReAct tasks.\n\n| Path | Policy | Purpose | \n|---|---|---|\n| `/inputs` | read-only | supplied documents and benchmark inputs | \n| `/workspace` | read-write | source checkout, extracted data, scratch work | \n| `/outputs` | controlled read-write | final persistent deliverables | \n\nFile and shell tools share this one task sandbox and path policy. Interactive\nsessions add an approval gate on writes, deletion, package installation, and\nrisky shell commands; some operations stay denied even with `--yes`; and file\nmutations are journaled so `/revert` can undo them.\n\nDetails: [sandboxing and path policy](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/framework.md#sandboxing),\n[approval and trace behavior](https://github.com/ApodexAI/FrontierAgent/blob/main/apodex/README.md#safety), and\n[the security policy](https://github.com/ApodexAI/FrontierAgent/blob/main/SECURITY.md).\n\n```\nuv sync --frozen --extra sandbox --extra document-readers --extra eval --extra dev\nuv run pytest -q\nuv run ruff check .\n```\n\nSee [CONTRIBUTING.md](https://github.com/ApodexAI/FrontierAgent/blob/main/CONTRIBUTING.md) for the full development environment,\npre-flight checks, session debugging, and submission process. Building the\ncontainer image is covered in\n[Run FrontierAgent in Docker](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/install/docker.md#build-from-the-current-checkout).\n\nThe evaluation harness runs each benchmark question in an isolated subprocess,\nsupports resumable multi-run experiments, and dispatches benchmark-specific\ndeterministic or model-based judges. A minimal smoke run, once the datasets are\ndownloaded per [the evaluation guide](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/eval.md#datasets), is:\n\n```\nuv sync --extra eval --extra sandbox --extra document-readers\nuv run python -m benchmarks.public.runner.run_subprocess \\\n  --benchmark browsecomp --pipeline stateful-react-agent --profile default \\\n  --limit 1 --concurrency 1 --out ./results/smoke\n```\n\nThe [evaluation guide](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/eval.md) is the canonical operator reference for\ncredentials, judge preflight, datasets, file benchmarks, execution, and result\ninspection. The [benchmark registry](https://github.com/ApodexAI/FrontierAgent/blob/main/benchmarks/README.md) lists dataset keys,\ndefault pipelines, scoring implementations, and extension points.\nFrontierSearchBench has its own external scorer and an isolation requirement, so\nit is documented separately in\n[FrontierSearchBench evaluation](https://github.com/ApodexAI/FrontierAgent/blob/main/docs/eval-frontier-search.md).\n\nBrowseComp, BrowseComp-ZH, xbench-DeepResearch, Humanity's Last Exam (text-only), SuperChem, FrontierScience-Research, FrontierScience-Olympiad, DeepSearchQA, WideSearch, FrontierSearchBench, OfficeQA, GDPval, APEX, and OneMillion-Bench.\n\nGDPval uses deterministic deliverable validation in this open-source harness;\nthe agentic pairwise grader is intentionally excluded. The\n[benchmark registry](https://github.com/ApodexAI/FrontierAgent/blob/main/benchmarks/README.md#supported-benchmarks) is authoritative\nfor each dataset key, its default pipeline, and its scoring implementation.\n\nThe chart above compares the two FrontierAgent workflows with the Apodex-1.0 baseline and selected external systems. The Apodex results are summarized here:\n\n| Configuration | APEX-Agents | GDPval | FrontierFinance | FrontierScience-Research | BioMysteryBench | HLE | \n|---|---|---|---|---|---|---|\n| Apodex-1.1 Agent Team | 38.5 | 78.8 | 54.3 | 63.3 | 35.3 | 56.1 | \n| Apodex-1.1 ReAct | 34.4 | 69.5 | 48.7 | 55.0 | 23.5 | 53.2 | \n| Apodex-1.0 | 16.5 | 59.3 | 40.3 | 28.3 | 17.6 | 49.0 | \n\nEarlier Apodex-1.0 checkpoints remain available in the\n[Hugging Face collection](https://huggingface.co/collections/apodex/apodex-1),\nwith model cards and serving guidance.\n\nCite the current release:\n\n```\n@article{apodex11,\n  title         = {Apodex-1.1: Scaling Agentic Intelligence for Complex Work},\n  author        = {Apodex Team},\n  year          = {2026},\n  eprint        = {2608.23283},\n  archivePrefix = {arXiv},\n  primaryClass  = {cs.AI},\n  url           = {https://arxiv.org/abs/2608.23283}\n}\n@misc{frontierchallenge,\n  title        = {FrontierChallenge: Evaluating Scientific Workflow Completion},\n  author       = {Su, Liangcai and Feng, Zhaopeng and Chen, Zhuo and Zhang, Zhen\n                  and Lin, Xiang and Li, Ruilin and Zhang, Handuo and Wang, Ning\n                  and Wen, Kailong and Guo, Yueqi and Xing, Feng and Guo, Yiling\n                  and Qian, Chenxiong and Du, Simon Shaolei and Bing, Lidong\n                  and Wang, Xinyu},\n  year         = {2026}\n}\n```\n\nFor work that refers specifically to the previous generation:\n\n```\n@techreport{apodex10,\n  title  = {Apodex-1.0: A Verification-Centric Agent Team for Discoverative Intelligence},\n  author = {Apodex Team},\n  year   = {2026}\n}\n```\n\nApache 2.0 — see [LICENSE](https://github.com/ApodexAI/FrontierAgent/blob/main/LICENSE).", "url": "https://wpnews.pro/news/apodex-frontieragent", "canonical_source": "https://github.com/ApodexAI/FrontierAgent", "published_at": "2026-09-16 11:00:38+00:00", "updated_at": "2026-09-16 11:13:30.936693+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-products", "developer-tools", "ai-research"], "entities": ["Apodex", "FrontierAgent", "Apodex-1.1", "ReAct", "Agent Team", "FrontierSearchBench", "FrontierChallenge", "AgentBus"], "alternates": {"html": "https://wpnews.pro/news/apodex-frontieragent", "markdown": "https://wpnews.pro/news/apodex-frontieragent.md", "text": "https://wpnews.pro/news/apodex-frontieragent.txt", "jsonld": "https://wpnews.pro/news/apodex-frontieragent.jsonld"}}