FrontierAgent

FrontierAgent is an agent runtime, terminal product, and evaluation suite open-sourced by ApodexAI, designed for long-horizon research and file-based work. Its command-line terminal interface (TUI) is called frontier-agent and includes two native workflows.

Purpose and Positioning FrontierAgent is designed to support agent tasks that require long-running operation and involve file reading and writing and command execution, such as research, information organization, and deliverable generation. It also serves as an evaluation suite for assessing the performance of Apodex models on a variety of benchmarks. The framework, tools, workflows, and evaluation layer remain separate to facilitate independent reuse.

Main Capabilities

1. ReAct workflow: a single stateful agent researches, reads files, writes deliverables, runs commands, and iterates within a task-level sandbox.
2. Agent Team workflow: a coordinator maintains a task board, dispatches independent work to parallel sub-agents, collects structured reports, and synthesizes results; an optional fast reporter is used for final evidence review.
3. Task board: the add_task and update_task events of Agent Team are displayed in real time in the TUI sidebar, with statuses including pending, active, completed, blocked, and cancelled.
4. Sandboxed file work: Shell and file tools share a task-level file system, with /inputs read-only, /workspace for working state, and /outputs for persistent deliverables; authorization and sandbox failures adopt a fail-closed policy.
5. Asynchronous intervention: entering new instructions while the agent is running queues them and injects them at the next safe turn boundary without discarding the current run; in Agent Team mode, it guides the coordinator, and already running sub-agents are allowed to complete.
6. Transparent deliverables: on macOS/Docker, /outputs maps to the host .apodex/runs/<session-id>/outputs, and the same run directory also contains checkpoints, traces, engine logs, and trajectories.
7. Approval, tracing, and recovery: mutating operations show diffs and require approval (unless --yes is enabled); sessions have checkpoints, every action is traced locally, /revert restores session changes, and --resume continues a saved run.
8. Built-in evaluation: a subprocess runner supports research and file-based benchmarks, deterministic artifact collection, concurrency, progress checks, and rerunning a single failed item.

Usage partiesFormat Requirements: Git, Python 3.12, uv, and an OpenAI-compatible model endpoint. Docker is optional. Installation: Install from the repository using uv; the endpoint can be saved in a user file. Exported variables and the project .env take precedence over the user file. Launch: Run frontier-agent to enter an interactive session, or pass in a task and stay in the session to follow up. Scripts: ./scripts/run-macos.sh and ./scripts/run-linux.sh set up a managed endpoint installation; ./scripts/run-linux-gpu.sh --install-system-deps --setup-only prepares a natively isolated SGLang environment on a Linux NVIDIA GPU. Docker: Prebuilt linux/amd64 and linux/arm64 images are published on the GitHub Container Registry, but the packages are private, requiring docker login ghcr.io or a local build. Local SGLang service: Pinned to a reviewed NVIDIA driver/CUDA/SGLang track; the driver must be confirmed against the GPU compatibility matrix.

Use cases

• Long-horizon research tasks that require the agent to autonomously read files, run commands, and produce deliverables.
• Complex tasks that require parallel sub-agent collaboration.
• Benchmarking Apodex models, supporting BrowseComp, BrowseComp-ZH, xbench-DeepResearch, Humanity's Last Exam (text-only), SuperChem, FrontierScience-Research, FrontierScience-Olympiad, DeepSearchQA, WideSearch, FrontierSearchBench, OfficeQA, GDPval, APEX, OneMillion-Bench, and more.
• Security-sensitive scenarios that require sandbox isolation, approval tracking, and session recovery.

Limitations

• Requires an OpenAI-compatible model endpoint; the Apodex API is optional (officially stated to be free for a limited two-week period; refer to the official source for specifics).
• The local SGLang service has compatibility requirements for NVIDIA driver/CUDA/SGLang versions, and a mismatch may cause CUDA or Triton kernel errors.
• The Docker image packages are private and require logging in with an authorized account.
• Parallel Search MCP does not support custom region, language, or time filtering, or more than 10 results per query; Serper should be used when these features are needed.
• GDPval uses deterministic deliverable verification in the open-source evaluation framework, intentionally excluding the agent pairwise grader.
• During evaluation, it is recommended to start with --concurrency 1; total concurrent model calls may approach the runner concurrency multiplied by the team generation limit.
• Some operations will still be rejected even with --yes enabled.

License: Apache 2.0.