FrontierAgent

FrontierAgent is Apodex's agent runtime, open-sourced alongside their model, and it aims at long-horizon file-based work rather than chat. The terminal UI ships two workflows. ReAct runs one stateful agent that researches, reads files, writes deliverables, and runs commands in a task-scoped sandbox. Agent Team puts a coordinator on a task board, farms independent work out to parallel sub-agents, collects structured reports, and synthesizes the result. Apache-2.0 and free.

The sandbox design is the part worth stealing. Shell and file tools share one task-scoped filesystem split three ways: /inputs is read-only, /workspace is working state, and /outputs holds the deliverables that persist. Authorization and sandbox failures are fail-closed, which is the correct default and not the common one. The task board renders live in the sidebar, and you can type while an agent is running to queue a new instruction instead of waiting for it to finish. Install is one command on macOS and Linux, no preinstall step and no hard Docker dependency.

The runtime costs nothing; model calls do. It speaks OpenAI-compatible endpoints, so point it at whatever you already pay for, or serve models locally through SGLang. Solo devs: free, and Agent Team is worth an evening if parallel sub-agents interest you. Teams: evaluate it, but read the catch first.

The catch is that this is a vendor's harness published to sell you on the vendor's model. The README leads with a limited-time free trial of the Apodex API and the benchmark chart is Apodex's own. Neither of those makes the code bad, and Apache-2.0 with OpenAI-compatible endpoints means you are not locked in. But the project is days old, the evaluation numbers are self-reported, and a framework this young will move under you. If you want a multi-agent runtime you can build on today, look at goose or the broader LangGraph and CrewAI ecosystem instead.