What it is
Open-source terminal agent runtime for long-horizon research and file work, with single-agent ReAct, parallel Agent Team mode, and a built-in evaluation harness. Run long-horizon research and file-based agent workflows from a native TUI or subprocess runner, with single-agent ReAct, parallel sub-agents, sandboxed files, approvals, resume, and benchmark evaluation.
Key results
- Two workflows: ReAct and Agent Team with bounded parallel sub-agents
- Task sandbox splits /inputs RO, /workspace RW, /outputs persistent
- Supports 13 benchmarks including BrowseComp, HLE, GDPval, and APEX
Shows that a reusable agent runtime can package execution, safety controls, and benchmark evaluation in one terminal-first workflow.