Conductor Java Agent SDK

Build durable Java AI agents on Conductor. Agents can use local Java tools, wait for people, execute dynamic plans, and recover after a process restart because Conductor persists execution state.

New here? Follow Getting Started to configure a server-side LLM provider and run the maintained basic-agent example.

For an AI coding agent that understands Conductor operations, load Conductor Skills: npm install -g @conductor-oss/conductor-skills && conductor-skills --all.

Install

Requirements: Java 21+ and a Conductor server.

dependencies {
    implementation 'org.conductoross:conductor-client-ai:<VERSION>'
}

Replace <VERSION> with a published version from Maven Central.

Start here

• Getting Started — configure a server and run a maintained agent example.
• Deploy · Serve · Run · Plan — choose the right runtime mode.
• Scheduling — schedule deployed agents with the shared typed client.

Build agents

• Agents — builder API and @AgentDef.
• Tools — Java tools, HTTP/MCP tools, human approval, files, and credentials.
• Multi-Agent — sequential, parallel, handoff, swarm, and plan-execute agents.
• Guardrails, Termination, Callbacks, Stateful Agents, Streaming & Human-in-the-Loop, and Structured Output.

Framework bridges

• Google ADK — bridge native ADK agents and sub-agent graphs.
• LangChain4j — turn existing @Tool POJOs and ChatModel metadata into Conductor agents.
• LangGraph4j — run a native graph builder on the durable runtime.
• OpenAI Agents SDK style — use familiar tool and handoff shapes in Java.

Operate and inspect

• Spring Boot — auto-configured runtime and @AgentDef discovery.
• Runtime reference and control-plane reference.
• API map, agent-definition fields, and configuration schema.
• Control-plane example — deploy, start, inspect, stop, and cancel an agent.

What Conductor adds

Capability	Conductor agent runtime
Process failure recovery	Durable workflow state resumes from completed work.
Java tools	Tools run as independently scalable Conductor worker tasks.
Long-running work	Human approval, schedules, and events do not hold application threads open.
Dynamic execution	Plans become durable sub-workflows that can be inspected and retried.
Observability	Inputs, outputs, tool calls, retries, and status share one execution record.