Getting started with Agent Eval

Agent Eval 0.18.0 is on Maven Central. Maven coordinates and Java packages still use agent-judge and io.github.markpollack.judge.

Use the least intelligent practical evaluator that establishes the property at the reliability you need. Keep the compiler, ordinary JUnit, and available deterministic analyzers for facts they can settle. Supply a bounded evidence bundle when a written requirement needs judgment.

Add the dependencies

You need Java 21. Most callers need the core API, the explicit Jackson 2 codecs, the model-backed requirements, and the fluent assertions:

<dependencies>
  <dependency>
    <groupId>io.github.markpollack</groupId>
    <artifactId>agent-judge-core</artifactId>
    <version>0.18.0</version>
  </dependency>
  <dependency>
    <groupId>io.github.markpollack</groupId>
    <artifactId>agent-judge-json-jackson2</artifactId>
    <version>0.18.0</version>
  </dependency>
  <dependency>
    <groupId>io.github.markpollack</groupId>
    <artifactId>agent-judge-ai-core</artifactId>
    <version>0.18.0</version>
  </dependency>
  <dependency>
    <groupId>io.github.markpollack</groupId>
    <artifactId>agent-judge-assertj</artifactId>
    <version>0.18.0</version>
    <scope>test</scope>
  </dependency>
</dependencies>

Add the runtime, execution and bridge modules you use from the inventory below. The requirements example has a complete caller POM and executable Java example with deterministic responses. For the PetClinic walkthrough, use the tutorial’s case-study run guide and presenter runbook; they cover IntelliJ, offline replay and caching. No live inference is required.

Build from source

To build the released source yourself, check out tag v0.18.0 of the Agent Judge repository and run ./mvnw -B -ntp clean install with Java 21. This installs the complete reactor, including the parent and every module, into your local Maven repository with tests enabled.

Configure, execute, retain

Define or read the actual Requirement and its immutable source/revision.
Select prepared evidence, an acquisition supplier, or a native investigative runtime configured with workspace/tools.
Construct a ready Judge or Jury using typed builders.
Execute judge() or vote(), or one cached AssertJ stage.
Keep the whole result and optionally request an application reliance policy.
Store/reopen V6 results with the explicit Jackson support module.

Dependencies

All artifacts retain group io.github.markpollack and are released at version 0.18.0. The AgentWorks BOM does not yet manage 0.18.0, so declare the version explicitly. The inventory is 14 JAR modules plus agent-judge-parent (POM packaging). The 0.18 migration guide explains the intentional API and wire breaks.

Artifact	Purpose
`agent-judge-core`	Pure requirements, Judge/Jury construction, results, voting, policy
`agent-judge-ai-core`	EvalModel, native RFC2119/EARS requirements and producers
`agent-judge-json-jackson2`	Explicit Jackson 2 converters and strict V6 codecs
`agent-judge-assertj` / `agent-judge-assertions`	Fluent assertions / retained assertion checks
`agent-judge-llm` / `agent-judge-agent-client`	Generated runtime adapters
`agent-judge-jev`	Typed structured native execution
`agent-judge-exec` / `agent-judge-file`	Build/command and file evaluation
`agent-judge-rag`	Typed retrieval evidence and RAG judges
`agent-judge-spring-ai` / `agent-judge-langchain4j` / `agent-judge-koog`	Evaluated response evidence bridges

The fundamentals path still demonstrates released 0.17.0 in its separate reactor. The PetClinic path demonstrates configured 0.18 APIs. For historical consumer changes, keep the 0.17 migration guide at its original semantics.

Projects

AgentWorks

Supporting Projects

Migration

Explorations

Older / Inactive Projects

Getting started with Agent Eval

Add the dependencies

Build from source

Configure, execute, retain

Dependencies

​Add the dependencies

​Build from source

​Configure, execute, retain

​Dependencies

Add the dependencies

Build from source

Configure, execute, retain

Dependencies