Build Agent Harnesses with JEV

This guide covers the dev version. Its APIs and examples may differ from released packages. For AgentScope Java 2.0.4, see the 2.0 documentation.

JEV supplies typed decisions where a Harness needs semantic judgment: which tools fit, whether an answer has evidence, which history to retain or when a task needs verification. Call components directly or integrate them into Agent execution as middleware or read-only tools.

Capabilities

Area	Typical task	Guides
Trace evaluation and reusable definitions	Check tool use, grounding and requirement coverage with shared offline and online metrics	Review definitions, trace evaluation, custom metrics, batch evaluation
Context management and recovery	Preserve constraints, reduce tool history and recover original material	Compaction and archives, explicit exchange planning, memory admission
Progress and delivery supervision	Observe drift, stalls, missing verification and readiness for review	Continuous supervision, verification evidence, checkpoint review
Tool selection and execution guards	Narrow a large catalog; check proposed refunds or outbound messages	Tool selection, pre-execution guards
Model, stage and team routing	Choose candidates by complexity, tools, capacity and quota	Per-call routing, explicit stages, team recommendations
Retrieval and knowledge answers	Filter injection, retain conflicting evidence, rerank results and check grounding	Evidence processing, relevance and grounding, Knowledge adapter
Content screening and answer revision	Screen input/output before publication; revise only rejected drafts	Content guardrails, Agent answer revision, standalone draft pipeline
Code review and evidence location	Screen diffs or source snapshots for risk and locate evidence	Staged code review
Read-only browser navigation and verification	Follow visible links to find information on authorized pages	Browser execution, action proposals only
Business classification and support	Route multiple requests and check conditional refunds and reply promises	Candidate classification, support review

Combine checks in a support workflow

For “refund me if it hasn’t shipped,” keep both lookup and refund tools during selection. Before dispatch, check that current evidence supports the refund. Review the final reply for invented payment or omitted conditions. After the invocation, use trace metrics to assess tool use and answer quality. Enable each capability independently. Start with the refund tool guard for context, configuration and failure handling, or run the offline example below to see a rejected draft revised without repeating an action.

Choose an integration

Direct API: If your application already has the input, evidence or draft, call the Judge, evaluator or evidence processor and handle the result.
Harness: Register middleware with .middleware(...), plug into .compaction(...), or register a Toolkit tool at the appropriate execution step.

A JEV decision grants no tool permission and does not prove a business operation succeeded. Your application owns authorization, ACLs, argument validation, idempotency and independent completion checks. Each guide distinguishes recommendations from delivered inputs and actual results.

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

Run the example to see retrieval, review and bounded revision without credentials. Runnable source is under agentscope-examples/jev in package io.agentscope.examples.jev. Before connecting a live model, read Noul, Choice, Score and decision statuses, then configure the client and execution mode. Calibrate example thresholds on your own samples.

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.