Tasks
A task packages instructions, an environment, and a verifier. It records the instructions in instruction.md and configuration in task.toml. Its environment provides the tools and files the agent needs; the verifier checks the resulting state.
View the task directory layoutTrials and results
A trial is one agent's attempt at a task. A job groups trials across tasks, agents, and models. The verifier scores the final state and writes reward.txt or reward.json. Its entry point is tests/test.sh, or test.bat on Windows.
Execution status and verifier reward describe different things. A completed execution alone does not establish that the task was solved; a missing reward remains unreported.
Agent trajectories
ATIF means Agent Trajectory Interchange Format. It represents recorded steps, including messages, tool calls and results, with agent and session metadata. Optional metrics and embedded subagent trajectories add context when present.
tune.new preserves recorded steps and shows the metrics available in the trace. Missing metrics remain unreported. Saved sessions preserve their recorded conversation without requiring an ATIF document.
tune.new is a SynthLabs product. Not affiliated with or endorsed by the Harbor team.
Back to documentation