conductor

Conductor is an open-source, durable execution workflow engine for the aforementioned microservices, AI agents, and adaptive workflow graphs. It was originally initiated by Netflix and is now actively maintained by Orkes and the community.

Purpose Conductor turns runtime choices—such as loops, branches, fan-out, tool calls, approvals, retries, and cancellation—into durable, inspectable execution processes for building reliable long-running workflows and AI agent orchestration. It makes the orchestration layer a declarative, machine-readable graph, while workers can be ordinary code in any language, any library, and any I/O, allowing developers to keep their preferred SDK or framework while operators still gain durable state, policy boundaries, replay, versioning, and auditability.

Key capabilities

• Durable execution and high resilience: Workflow and task state are persisted, supporting recovery after worker and infrastructure failures, and providing retries, timeouts, pause, resume, and termination controls.
• Control flow constructs: Supports SWITCH (conditional branching), DO_WHILE (loops with configurable iteration cleanup), FORK_JOIN (parallel execution and dynamic fan-out), SUB_WORKFLOW (composition), and DYNAMIC tasks resolved at runtime, and these can be nested within each other; each task supports configurable retries, timeouts, and optional/compensation tasks.
• Versioning: Workflow definitions are versioned by number, running executions continue to use the version from when they started, and deploying a new version does not break in-progress workflows; there is no need to worry about code replay compatibility.
• Native AI capabilities: Provides native LLM tasks, MCP tool discovery and invocation, human approval, and vector workflows for RAG; supports 14+ LLM providers, function calling, and vector database integrations.
• Observability and validation: Built-in visual UI for designing, running, and debugging workflows; every execution is fully observable, and the input, output, duration, and retry history of each task can be inspected; workflow inputs and task I/O are validated via JSON Schema.
• Long-running: Execution is fully durable, workflows can pause for months waiting for human approval, external signals, or timers, and resume exactly from where they left off; AI agent loops can continue after waiting for human review.
• Runtime flexibility: Workflows are JSON, and LLMs can generate and modify them at runtime without a compile/deploy cycle; supports dynamic fan-out to a runtime-determined number of parallel tasks, as well as dynamic sub-workflows that compose other workflows by name; workers are decoupled from orchestration, allowingindependent updating of the workflow graph or replacement of worker implementations.

How to use Prerequisites for quick start: Node.js v16+ and Java 21+ are installed. After starting, open http://localhost:8080，服务器会运行并带有内置 ui-next UI. If upgrading from an older version, the CLI caches the server JAR in ~/.conductor-cli/; if an older version is cached, you need to force a re-download. You can run Conductor (including the ui-next UI) using a Docker image. All CLI commands have equivalent cURL/API calls. For example command runs, refer to the Quickstart guide, which includes complete steps for writing a worker and replaying a workflow; note that if you run the same command twice, the second time returns an error because the workflow already exists. This is expected behavior, and you can use conductor workflow update to modify an existing workflow.

AI Coding Assistant Skills Conductor Skills allow AI coding assistants (such as Claude Code, Gemini CLI, etc.) to create, manage, and deploy Conductor workflows directly in the terminal. A single command can automatically detect all supported agents in the system and install globally where possible; it can be re-run at any time, and it will only install for newly detected agents. Windows provides both PowerShell and cmd methods.

Building from Source Requirements: Docker Desktop, Java (JDK) 21+, Node.js 18+, and pnpm (for the UI). Run the UI in development mode (hot reload at http://localhost:1234）需要 The Conductor server runs at http://localhost:8080；如未启用 corepack needs to be enabled once first, then start the development server and open http://localhost:1234，文件更改后 The UI reloads automatically. The complete build guide can be found in the documentation.

Use Cases

• Build autonomous think-act agents: discover tools via MCP, reason with an LLM, invoke the chosen tools, and repeat until done.
• Governed adaptive graph patterns: plan → verify approved capabilities → bounded fan-out or human approval → evaluate → continue or end.
• Production workflows that require long waits for human approval or external signals.
• AI applications that need to orchestrate multiple LLM providers, MCP tool calls, and RAG vector workflows.

Limitations and Notes

• Execution history retention depends on configuration; keepLastN intentionally removes older loop iterations.
• This material does not provide specific fees, pricing, or SaaS service details; however, it mentions Orkes as the primary maintainer and provides an enterprise-grade SaaS platform across major cloud providers.
• This open-source version is licensed under Apache 2.0, is self-hostable, and supports 5 persistence backends and 6 message brokers; the Conductor server and workers can scale independently, and task domains, concurrency limits, persistence configuration, and metrics can be used to match throughput and isolation requirements.
• Orkes Conductor is built on Conductor OSS, with fully compatible APIs and workflows.