AI NewsOpen sourceAnnouncement
ApodexAI open-sourced FrontierAgent, which reached 1,360 stars in 10 days
ApodexAI released FrontierAgent, an Apache-2.0 Python agent framework with a terminal UI and a coordinator-plus-subagents mode, alongside its Apodex-1.1 model.
Image: GitHub
Why it mattersA working developer weighing whether to build a research agent in-house now has one more open framework to try before writing the connecting code themselves.
ApodexAI has open-sourced FrontierAgent, a Python agent runtime published under Apache-2.0 on GitHub. The repository was created on 22 August and has gained 1,360 stars and 125 forks in the ten days since, one of the fastest-growing repositories on the GitHub trending list this week.
The framework comes with a native terminal user interface, an evaluation suite, and two agent modes. It is released alongside Apodex-1.1, the lab's own model, but the framework itself is model-agnostic and works with any OpenAI-compatible endpoint.
The two modes
In ReAct mode, a single stateful agent researches, reads and writes files, runs shell commands, and iterates inside a task-scoped sandbox. In Agent Team mode, a coordinator keeps a task board, delegates independent pieces of work to parallel sub-agents, collects their reports, and synthesises the result. The README describes the two as different trade-offs between control and parallel work, with ReAct better for tightly coupled reasoning and Agent Team better for work that splits easily into separate parts.
The runtime, tools, workflows, and evaluation layer are kept separate on purpose, so a team can swap one part without rewriting the others.
The benchmark numbers
ApodexAI reports Apodex-1.1 in Agent Team mode scoring 78.8 on GDPval, 63.3 on FrontierScience-Research, 56.1 on Humanity's Last Exam, and 38.5 on APEX-Agents. It reports the same model in ReAct mode at 69.5, 55.0, 53.2, and 34.4 on the same four benchmarks, and Apodex-1.0 at 59.3, 28.3, 49.0, and 16.5.
Every one of those numbers is a self-report against the lab's own baseline model. The README leaves independent comparisons against other leading models for the reader to run. Treat the figures as a claim from the vendor until someone else measures them.
Requirements to try it
FrontierAgent needs Python 3.12, uv, and an OpenAI-compatible model endpoint. Docker is optional. The README lists support for local SGLang serving on NVIDIA GPUs for teams that want to run models on their own hardware.
This joins many similar tools, including AutoGen, LangGraph, CrewAI, and the coding-agent CLIs from OpenAI and Anthropic. What stands out is that ApodexAI has separated the framework from its model, published both under permissive licences, and included an evaluation harness in the same repository as the code. A team deciding whether to write agent infrastructure code in-house can read this one end to end without licence concerns.
More usefully, the same eval harness runs against any OpenAI-compatible endpoint. That means a team can run it against GPT, Claude, Gemini, or a local model and see the number for themselves. On a framework release whose selling point is a large jump on self-reported benchmarks, that check is the one that matters.
The star count says other developers are looking. Whether the Agent Team mode has a real advantage over a stronger leading model in ReAct is the next thing to find out, and the harness makes it a one-command test.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
