FrontierAgent: a new agent environment with TUI, Agent Team mode and quality evaluation

FrontierAgent is an open agent environment for long-running research and file tasks; the repository was opened on August 22, 2026, and in 28 days it gained about 4 thousand stars. It is a runtime, a terminal product and an agent evaluation kit in one project: two native operating modes — ReAct and Agent Team — launch with a single command on macOS and Linux without mandatory Docker.

What it does

• Agent Team — a coordinator maintains a task board, hands out independent work to parallel subagents, collects their reports and synthesizes the result.
• ReAct mode — a single stateful agent explores, reads files, writes results, executes commands and iterates in a sandbox.
• Sandbox for files and commands — shell and file tools run in the task's shared file system: /inputs is read-only, /workspace is working state, /outputs is persistent results; authorization and sandbox failures are closed off by denial.
• Human confirmation — mutating operations show a diff and require approval until --yes is enabled; sessions are checkpointed, every action is traced locally, /revert rolls back changes, --resume continues a saved run.
• Asynchronous intervention — you can type while the agent works: the instruction is queued and injected at the nearest safe turn boundary without discarding the active run.
• Agent quality evaluation — the evaluation harness runs each benchmark question in an isolated subprocess, supports resumable repeated experiments and deterministic or model-based judges.

Your own model instead of the cloud

Your own model is connected via an OpenAI-compatible API and SGLang. The documentation states that local SGLang serving is pinned to validated tracks of NVIDIA, CUDA and SGLang drivers; model hosting is not required — an API key and a compatible endpoint are enough. The platform is written in Python, installs via docker run, license Apache-2.0.

How it differs from its neighbors

Of the three neighbors, only FrontierAgent claims graphs and state persistence in its documentation, as well as agent quality evaluation. At the same time, MCP support, agent memory and its own server with Docker are mentioned by most neighbors, while here they are not claimed in the documentation. Custom tools and functions are claimed by all four platforms. In terms of connecting your own model, FrontierAgent is closer to DeerFlow and OpenBot via an OpenAI-compatible API, whereas Hermes Agent bets on Ollama, llama.cpp and vLLM.

How alive the project is

The repository is less than a month old, with 10 contributors and about 4 thousand stars. This is an early project with an active start, but without a long release history — for comparison, Hermes Agent has 16 stable releases in 90 days, and OpenBot has 12.

FrontierAgent will suit those who build agents for long-running research and file tasks, want built-in quality evaluation and are ready to work with an OpenAI-compatible API or SGLang. You can follow new platforms and breakdowns in the CPU3D news feed.