Run the JEV examples - AgentScope Java

All examples are in agentscope-examples/jev, in package io.agentscope.examples.jev. Start with synthetic decisions to exercise the Harness, then substitute your own backend and business evidence.

Build once and run multiple examples

From the repository root:

mvn -pl agentscope-dependencies-bom,agentscope-distribution/agentscope-bom,agentscope-examples/jev -am install -DskipTests
mvn -pl agentscope-examples/jev dependency:build-classpath -Dmdep.outputFile=/tmp/jev-examples-cp.txt
JEV_CP="agentscope-examples/jev/target/classes:$(cat /tmp/jev-examples-cp.txt)"
java -cp "$JEV_CP" io.agentscope.examples.jev.JevIntegrationExample

The combined example looks up refund conditions, retrieves authorized evidence, generates a draft that omits a condition, then revises only the text after review fails. The final answer retains the refund conditions. It uses in-memory retrieval, a scripted generation model and synthetic decisions, without credentials or model service calls.

What to expect

The final reply should retain both conditions: the item must be unused and within the seven-day window. The generation model runs three times: propose retrieval, generate an incomplete draft, then rewrite text without tools. Console review statistics confirm the scripted branches rather than online accuracy. To explore one capability, run its standalone scenario below and use the corresponding API guide to replace the inputs and backend.

Run examples by capability

Replace the class name at the end of the previous Java command with an entry below. All run offline by default, except for the local browser example described separately.

Capability	Entry point and full source
Standalone tool selection and model routing	JevHarnessScenarios, with `selection` or `routing`
Tool selection, guards and per-call routing	JevHarnessExample
Customer support Judge	JevCustomerSupportExample
Retrieval, guardrails and answer revision	JevIntegrationExample
Trace evaluation	JevTraceEvaluationExample
Context compaction and recovery	JevContextCompactionExample
Continuous supervision and verification evidence	JevSupervisionExample
Code review	JevCodeReviewExample
Retrieved evidence and draft revision	JevEvidenceExample
Automatic and explicit stage routing	JevPhaseRoutingExample
Read-only browser tool	JevBrowserExample
Standalone application and review scenarios	JevApplicationScenarios; omit arguments to run all, or specify `classification`, `context`, `memory`, `support`, `team`, `browser`, `judge`, `evaluation`, `task-review`, `rag` or `draft`
Classification, memory, teams and application components	JevApplicationExample

Scripted probabilities demonstrate status and execution branches; they cannot measure real model accuracy. For benchmark inputs and live backend options, see Evaluate your Harness.

Use a local browser

JevBrowserLocalExample serves a loopback page and uses an explicitly supplied Chromium executable. By default it uses synthetic decisions and does not access external websites or model APIs:

java -cp "$JEV_CP" io.agentscope.examples.jev.JevBrowserLocalExample \
  --browser "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"

Substitute the browser path for your system. The session visits the policy page, an independent verifier checks its content, and the page and local service close. Browser API

Make live calls

The support example calls the client only with --live. Other ordinary examples have no common live switch: replace the scripted caller and model as shown in each guide. Browser and benchmark programs have their own explicit --live-jev / --live-qwen options. Environment variables alone never switch an offline example to live calls.