Skip to main content
The core evaluation flow in Halios follows a simple progression:
A scenario is the reusable test. A trace is the execution evidence from one trial. Running the scenario again produces a fresh trace against the current version of your agent.

Core terms

Scenario

The reusable test situation you want to evaluate. It defines the initial user goal, persona, constraints, and multi-turn conversation arc.

Trial

One fresh execution generated from a scenario. Instead of replaying a saved transcript, Halios simulates a live user interaction with your agent for each trial.

Trace

The execution evidence emitted by your agent during a trial. Contains the full conversational turn history, model calls, tool arguments, retrieval context, and latency metrics.

Check

One specific invariant you want to verify about your agent’s behavior (e.g. answer groundedness, regex format match, policy compliance, or tool argument schema).

Eval suite

The collection of scenarios, checks, and configuration defined for an agent or workflow (stored in your repository under .halios/).

Run

One complete execution pass of an eval suite against the current agent implementation, recording trial traces, check evaluations, and pass rates.

pass@k

Repeating the same scenario kk times (e.g. 3 or 5 trials) to measure consistency. A scenario passes only if all kk trials succeed, ensuring that passing was not just random chance.

Run your first eval

Get started with your coding agent in minutes.

CLI reference

Command reference for automation and CI pipelines.