AgentOptimizationOptions interface

Package:

@azure/ai-projects

Tuning knobs and run-mode for an optimization job.

Properties

evaluation_level	Evaluation granularity. Null/omitted means per-item single-turn. Set to 'conversation' for per-conversation multi-turn simulation scoring.
eval_model	Model deployment used for evaluation. Defaults to server config (typically 'gpt-4o').
max_candidates	Maximum number of optimization candidates to generate. Must be >= 1. Default: 5.
max_stalls	Maximum number of consecutive reflective minibatch rejections before stopping early. A 'stall' occurs when the optimizer proposes a prompt change, evaluates it on a small subset, and the score does not improve — so no full validation-set evaluation is triggered. The counter resets whenever a minibatch passes and its full-validation score beats the current best. Only a sustained plateau of max_stalls consecutive minibatch failures triggers the stop. The service defaults to 5 if a value is not specified by the caller. Must be >= 1 when set.
optimization_config	Per-target-attribute configuration overrides. Contains skills, tools, system_prompt for the agent, plus model space for model optimization.
optimization_model	Model deployment for optimization reasoning (must be gpt-5 family). Falls back to the default eval model when not set.

Property Details

evaluation_level

Evaluation granularity. Null/omitted means per-item single-turn. Set to 'conversation' for per-conversation multi-turn simulation scoring.

evaluation_level?: EvaluationLevel

Property Value

EvaluationLevel

eval_model

Model deployment used for evaluation. Defaults to server config (typically 'gpt-4o').

eval_model?: string

Property Value

string

max_candidates

Maximum number of optimization candidates to generate. Must be >= 1. Default: 5.

max_candidates?: number

Property Value

number

max_stalls

Maximum number of consecutive reflective minibatch rejections before stopping early. A 'stall' occurs when the optimizer proposes a prompt change, evaluates it on a small subset, and the score does not improve — so no full validation-set evaluation is triggered. The counter resets whenever a minibatch passes and its full-validation score beats the current best. Only a sustained plateau of max_stalls consecutive minibatch failures triggers the stop. The service defaults to 5 if a value is not specified by the caller. Must be >= 1 when set.

max_stalls?: number

Property Value

number

optimization_config

Per-target-attribute configuration overrides. Contains skills, tools, system_prompt for the agent, plus model space for model optimization.

optimization_config?: Record<string, any>

Property Value

Record<string, any>

optimization_model

Model deployment for optimization reasoning (must be gpt-5 family). Falls back to the default eval model when not set.

optimization_model?: string

Property Value

string