Zenith

Zenith targets work that may run for days or weeks, where the researchers behind it argue the main failure mode is an agent stopping before a task is truly complete rather than being unable to make progress. It runs one orchestrator session over MCP and the Agent Client Protocol that reads the current task state each turn and decides whether to spawn worker and tester agents, register a reusable skill, revise the plan, or stop.

The accompanying technical report compares five harness designs across eight long-horizon tasks and finds that a simpler baseline called RALPH, which repeatedly reopens the gap between current state and the original requirement, is a strong but expensive approach with no principled stopping rule; Zenith is presented as keeping that benefit while adaptively allocating workers, testers and replanning, reportedly at under half of RALPH's per-task cost in the study. Installation uses uv and Node.js, with ACP adapters installed separately for whichever agents Zenith will run as workers.