Build Agent Harnesses with JEV

JEV adds typed semantic decisions to your Harness: tool selection, response review, context retention, routing and task supervision. Use components directly, register middleware or tools, or configure supported capabilities through AgentScope Service. For example, a support agent can select order tools, check a conditional refund before dispatch, review its final reply against policy evidence, and evaluate the completed trace. Enable each step independently; a semantic approval never proves that the refund was executed.

Capabilities

Area	User guide
Agent trace evaluation and reusable definitions	Trace evaluation, definitions
Context governance and recovery	Compaction, archives, memory admission
Long-running task supervision	Progress observation, verification evidence
Tool selection and execution checks	Selection, pre-execution checks
Model, stage and team routing	Model routing, stages, team recommendations
Retrieval evidence and answer quality	Evidence processing, content checks, draft revision
Code review and browser tasks	Code evidence, read-only navigation
Business classification and support	Candidate selection, support review

Detailed guides are currently available in Chinese. Each explains the supported behavior, application inputs and APIs.

Start with an offline example

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
java -cp "agentscope-examples/jev/target/classes:$(cat /tmp/jev-examples-cp.txt)" io.agentscope.examples.jev.JevIntegrationExample

This example uses scripted generation and judgment responses; no API key is required. The final reply retains the unused-order and seven-day conditions. The example checks that revision uses no tools. See the scenario catalog for focused examples and their API guides. Runnable examples live in agentscope-examples/jev. Add agentscope-extensions-jev to your application and the same-version agentscope-harness dependency when using Harness integrations.

Adopt decisions explicitly

Middleware defaults to OFF. SHADOW records recommendations while preserving downstream input; ENFORCE applies the component-specific policy. Trace evaluation and ongoing supervision support only OFF/SHADOW. Direct Judge calls evaluate when invoked. Semantic decisions do not grant permissions or prove execution succeeded. Keep deterministic authorization, ACLs and completion checks in your application. Calibrate thresholds using your own labeled data. See client setup, runtime modes, evaluation and Service configuration. When tool selection is enabled, JevToolSelectionMiddleware defaults to ContextStrategy.RECENT_WINDOW: the latest 8 nonblank text messages, with a total text budget of 8000 characters. Configure maxContextMessages and maxContextChars with positive values, use LATEST_USER_MESSAGE for only the latest user text, or FULL_CONVERSATION for the complete original message list. These options preserve the default OFF mode and independent per-tool relevance scoring.

​Capabilities

​Start with an offline example

​Adopt decisions explicitly

Capabilities

Start with an offline example

Adopt decisions explicitly