uip agent eval

uip agent eval is the evaluation command group. It manages evaluators, evaluation sets, and test cases on disk, and executes evaluation runs against the Agent Runtime service. Four subcommand trees live underneath:

• uip agent eval evaluator … — manage evaluators (evals/evaluators/*.json).
• uip agent eval set … — manage evaluation sets (evals/eval-sets/*.json).
• uip agent eval add | list | remove — manage test cases (evaluations) inside an evaluation set.
• uip agent eval run … — start, monitor, list, and compare evaluation runs against the Agent Runtime service.

Everything except uip agent eval run * is local-only. The run subcommands require an active CLI session (uip login) and that the agent has already been pushed to Studio Web (via uip agent push) or that you pass --solution-id explicitly.

All uip agent eval subcommands honor the global options (--output, --output-filter, --log-level, --log-file). Exit codes follow the standard contract.

Synopsis

uip agent eval evaluator add    <name> --type <type> [--description <d>] [--prompt <p>] [--target-key <k>] [--path <dir>]
uip agent eval evaluator list                                                                                 [--path <dir>]
uip agent eval evaluator remove <id>                                                                           [--path <dir>]

uip agent eval set add    <name>  [--evaluators <ids>]                                                         [--path <dir>]
uip agent eval set list                                                                                         [--path <dir>]
uip agent eval set remove <id>                                                                                  [--path <dir>]

uip agent eval add    <name> --set <name> --inputs <json>
                              [--expected <json>] [--expected-agent-behavior <text>]
                              [--simulation-instructions <text>]
                              [--simulate-input] [--simulate-tools]
                              [--input-generation-instructions <text>]
                              [--path <dir>]
uip agent eval list   --set <name>                                                                              [--path <dir>]
uip agent eval remove <id> --set <name>                                                                         [--path <dir>]

uip agent eval run start   --set <name> [--solution-id <id>] [--wait] [--timeout <s>]                           [--path <dir>]
uip agent eval run status  <evalSetRunId> --set <name>                                                          [--path <dir>]
uip agent eval run results <evalSetRunId> --set <name> [--only-failed] [--verbose] [--export-format <json|csv>] [--path <dir>]
uip agent eval run list    --set <name>                                                                         [--path <dir>]
uip agent eval run compare <evalSetRunId> --compare-to <id> --set <name>                                        [--path <dir>]

uip agent eval evaluator

Manage evaluators — the graders that score an agent's output.

eval evaluator add

Arguments:

• <name> (required) — Evaluator name. Used as the default file name.

Options:

Drapeau	Default	Requis	Objectif
--type <type>	—	oui	Evaluator type. Known values include SemanticSimilarity, Trajectory, and LLM-as-judge; run uip agent eval evaluator add --help for the full list supported by your installation.
--description <desc>	—		Free-text description.
--prompt <prompt>	—		Custom evaluation prompt (for prompt-based evaluators).
--target-key <key>	—		Target output key to evaluate against.
--path <path>	.		Path to the agent project directory.

Exemple :

uip agent eval evaluator add content-check \
  --type SemanticSimilarity \
  --path ./my-agent

Data shape (--output json):

{
  "Code": "AgentEvaluatorAdd",
  "Data": {
    "Status": "Evaluator added",
    "Name": "content-check",
    "Type": "SemanticSimilarity",
    "Id": "a1b2c3d4-0000-0000-0000-000000000130",
    "File": "content-check.json"
  }
}

eval evaluator list

Options: --path <path> (default .).

Data shape:

{
  "Code": "AgentEvaluatorList",
  "Data": [
    { "Name": "content-check", "Type": "SemanticSimilarity", "Id": "…", "File": "content-check.json" }
  ]
}

Empty projects return Data: { "Message": "No evaluators configured" }.

eval evaluator remove

Arguments: <id> — evaluator ID or name.

Options: --path <path> (default .).

Data shape:

{ "Code": "AgentEvaluatorRemove", "Data": { "Status": "Evaluator removed", "Id": "content-check" } }

uip agent eval set

Manage evaluation sets — named collections of test cases plus the evaluators that should score them.

eval set add

Arguments: <name> — evaluation-set name.

Options:

Drapeau	Default	Objectif
--evaluators <ids>	all evaluators in the project	Comma-separated evaluator IDs to include.
--path <path>	.	Path to the agent project directory.

Exemple :

uip agent eval set add smoke-tests \
  --evaluators a1b2c3d4-0000-0000-0000-000000000130,a1b2c3d4-0000-0000-0000-000000000131 \
  --path ./my-agent

Data shape:

{
  "Code": "AgentEvalSetAdd",
  "Data": {
    "Status": "Evaluation set created",
    "Name": "smoke-tests",
    "Id": "a1b2c3d4-0000-0000-0000-000000000110",
    "Evaluators": 2
  }
}

eval set list

Options: --path <path> (default .).

Data shape:

{
  "Code": "AgentEvalSetList",
  "Data": [
    { "Name": "smoke-tests", "Id": "…", "Evaluations": 5, "Evaluators": 2 }
  ]
}

eval set remove

Arguments: <id> — eval-set ID or name.

Options: --path <path> (default .).

uip agent eval add | list | remove (test cases)

Manage the test cases (evaluations) inside a set. These subcommands sit directly under eval, not under eval set.

eval add

Arguments: <name> — test-case name.

Options:

Drapeau	Default	Requis	Objectif
--set <name>	—	oui	Evaluation set name or ID.
--inputs <json>	—	oui	Input values as a JSON string. Parsed; invalid JSON fails fast.
--expected <json>	—		Expected output as JSON.
--expected-agent-behavior <text>	—		Expected behaviour description for trajectory evaluators (for example, "Must call Web Search tool").
--simulation-instructions <text>	—		Instructions for simulating agent behaviour during evaluation.
--simulate-input	off		Enable input simulation for this test case.
--simulate-tools	off		Enable tool simulation for this test case.
--input-generation-instructions <text>	—		Instructions for synthesizing inputs.
--path <path>	.		Path to the agent project directory.

Exemple :

uip agent eval add simple-greeting \
  --set default \
  --inputs '{"input":"hello"}' \
  --expected '{"content":"world"}' \
  --path ./my-agent

Data shape:

{
  "Code": "AgentEvalAdd",
  "Data": {
    "Status": "Evaluation added",
    "Name": "simple-greeting",
    "Id": "a1b2c3d4-0000-0000-0000-000000000120",
    "Set": "default"
  }
}

eval list

Options: --set <name> (required), --path <path> (default .).

Data shape:

{
  "Code": "AgentEvalList",
  "Data": [
    {
      "Name": "simple-greeting",
      "Id": "…",
      "Inputs": "{\"input\":\"hello\"}",
      "Expected": "{\"content\":\"world\"}",
      "ExpectedBehavior": "-"
    }
  ]
}

eval remove

Arguments: <id> — evaluation ID or name.

Options: --set <name> (required), --path <path> (default .).

uip agent eval run

Execute, monitor, and compare evaluation runs via the Agent Runtime service (EvalsTenantExecutionApi). Requires uip login.

eval run start

Start an evaluation run. The agent must already be in Studio Web (uip agent push) — either pass --solution-id explicitly or rely on SolutionStorage.json, which push writes automatically.

Options:

Drapeau	Default	Requis	Objectif
--set <name>	—	oui	Evaluation set name or ID.
--solution-id <id>	De SolutionStorage.json		Cloud solution ID. If omitted, the command reads SolutionStorage.json from the project; if neither is available, it errors out.
--path <path>	.		Path to the agent project directory.
--wait	off		Poll until the run completes and then emit summary + per-test-case rows.
--timeout <seconds>	600		Maximum seconds to poll when --wait is set.

Exemple :

uip agent eval run start --set default --path ./my-agent --wait

Data shape — kickoff (Code: "AgentEvalRunStarted"):

{
  "Code": "AgentEvalRunStarted",
  "Data": {
    "EvalSetRunId": "a1b2c3d4-0000-0000-0000-000000000101",
    "EvalSetName": "default",
    "TestCases": 5,
    "Evaluators": 2
  }
}

With --wait, two additional payloads follow after polling:

• Code: "AgentEvalRunCompleted" — summary (Status, Score, Duration, EvaluatorScores, TestCases).
• Code: "AgentEvalRunResults" — per-test-case rows (same shape as eval run results).

eval run status

Poll the status of an in-flight or finished run.

Arguments: <evalSetRunId> — run ID from eval run start.

Options: --set <name> (required), --path <path> (default .).

Data shape:

{
  "Code": "AgentEvalRunStatus",
  "Data": {
    "EvalSetRunId": "…",
    "Status": "completed",
    "Score": 0.86,
    "Duration": "42.5s",
    "EvaluatorScores": "semantic: 0.9, trajectory: 0.82"
  }
}

eval run results

Fetch per-test-case results.

Arguments: <evalSetRunId>.

Options:

Drapeau	Default	Requis	Objectif
--set <name>	—	oui	Evaluation set name or ID.
--path <path>	.		Path to the agent project directory.
--only-failed	off		Show only failed or errored test cases.
--verbose	off		Include evaluator justifications in the output.
--export-format <json|csv>	—		Write the formatted rows to eval-results-<timestamp>.(json|csv) instead of printing them.

Exemple :

uip agent eval run results <evalSetRunId> --set default --verbose --only-failed

Data shape (inline — no export):

{
  "Code": "AgentEvalRunResults",
  "Data": [
    {
      "TestCase": "simple-greeting",
      "Status": "completed",
      "Score": 1,
      "EvaluatorScores": "semantic: 0.95",
      "Tokens": 320,
      "Duration": "1.8s",
      "Error": "-"
    }
  ]
}

When --export-format is set, the payload becomes Code: "AgentEvalRunExported" with Format, File, and Records.

eval run list

List all runs for a given eval set.

Options: --set <name> (required), --path <path> (default .).

Data shape:

{
  "Code": "AgentEvalRunList",
  "Data": [
    {
      "EvalSetRunId": "…",
      "Status": "completed",
      "Score": 0.86,
      "TestCases": 5,
      "Duration": "42.5s",
      "EvaluatorScores": "semantic: 0.9, trajectory: 0.82",
      "CreatedAt": "2025-04-15T10:30:00Z"
    }
  ]
}

eval run compare

Compare two runs side by side. Useful for A/B testing prompt or model changes.

Arguments: <evalSetRunId> — first (baseline) run ID.

Options:

Drapeau	Default	Requis	Objectif
--compare-to <id>	—	oui	Second run ID to compare against.
--set <name>	—	oui	Evaluation set name or ID.
--path <path>	.		Path to the agent project directory.

Data shape (Code: "AgentEvalRunComparison"):

{
  "Code": "AgentEvalRunComparison",
  "Data": {
    "RunA": { "Id": "…", "Score": 0.86, "Status": "completed" },
    "RunB": { "Id": "…", "Score": 0.80, "Status": "completed" },
    "ScoreDelta": 0.06,
    "TestCases": [
      { "TestCase": "simple-greeting", "ScoreA": 1, "ScoreB": 0.9, "Delta": "+0.1", "StatusA": "completed", "StatusB": "completed" }
    ]
  }
}

Related

• uip agent push — must be run before eval run start (unless --solution-id is supplied).
• uip agent validate — the default eval set and evaluators are created by init; validate keeps them consistent.
• uip agent run — run the agent as an Orchestrator job; distinct from an Agent Runtime eval run.

Voir également

• Authentication — sessions and token validity for the eval run subcommands.
• Global options, Exit codes.