AI system that takes action and makes decisions

• Target Agent The AI agent you want to test
• Maihem Agent An AI agent that simulates interactions with your Target Agent

A sequence of workflow steps

• Workflow step An operation with determined input and output formats

The context in which the codebase is executed (production, dev, local, etc.)

• Revision A unique label for the state of the codebase

A particular realized sequence of transactions between a target agent and a user

• Conversation A sequence of messages between a target agent and a user

A corpus of text sent by the user or the target agent When a target agent generates a message, Maihem collects:

• Trace A particular sequence of spans
• Span A particular realization of a workflow step

Quantitative measure used for tracking and comparing performance

• Criteria Statement that can be falisified with certainty – used to flag failures

Connector between workflow steps and metrics, that maps required inputs and outputs

A qualitative judgement on a criteria using a metric. It contains:

• Score A numerical value against metric
• Is failed A boolean value with a pass/fail judgement
• Explanation String with details behind the judgement

A configuration of a procedure, used to evaluate a workflow or workflow step of a target agent. It maps to a specific evaluator Can be conducted with:

• Dataset Uploaded data with inputs for the workflow step, and optional expected outputs (ground truth)
• Maihem agents Simulated and dynamic inputs for the workflow step, and optional expected outputs (ground truth)

A particular execution of a test, used to ensure and compare quality levels among different versions and environments of the target agent. It contains:

• Interactions
• Evaluations
• Detected failures