Conductor for AI Assistants
What Conductor is
Conductor is an open-source durable execution platform for workflows, adaptive agents, and AI systems. A workflow is a versioned graph of tasks. Conductor persists execution state and coordinates task scheduling; workers and built-in system tasks perform the work.
Conductor supports two complementary AI paths:
- Native AI workflows: compose LLM, MCP, vector, human approval, and control-flow system tasks in a workflow definition.
- Framework-authored agents: compile a supported SDK or framework agent—such as OpenAI Agents, Google ADK, LangChain, or LangGraph—into a Conductor graph, then use it in a larger workflow.
Use the Agents & AI overview for the product map and framework agent recipes for supported bridges.
Safe authoring rules
- Prefer a built-in system task when it matches the operation. Do not replace native LLM, MCP, vector, approval, wait, transform, or control-flow tasks with an HTTP wrapper or a custom worker.
- Every external side effect must be idempotent. Conductor task delivery is at least once, so a task can be redelivered after failure or timeout.
- Bound adaptive execution. Use loop iteration caps, task and workflow timeouts, bounded fan-out, and approved capability selection.
- Do not put credentials in workflow input or prompts. Use the appropriate server-side integration, secret facility, or worker environment instead.
- Treat a generated workflow definition as untrusted data. Validate its structure and capability allowlist before starting it with
workflowDef. - Require approval before consequential writes. Use
HUMANdirectly or the SDK agent tool approval configuration. - Keep outputs intentionally small. Store large objects externally and pass references through the workflow.
Choose the right starting point
| Goal | Start here |
|---|---|
| Create a durable service workflow | First workflow |
| Build a governed plan/act/evaluate loop | Durable Adaptive Graphs |
| Bring an existing framework agent (LangChain, ADK, and more) | Framework Agent Bridges |
| Add policy and approval | Agent Guardrails |
| Test routes, tools, and output quality | Agent Evals |
| Design a production agent system | Production Agent Architecture |
| Check task and API fields | Workflow definition reference |
Durable execution vocabulary
- Workflow definition: versioned task graph; a running workflow uses the definition version it started with.
- Task: a unit of work. Built-in system tasks run in the platform;
SIMPLEtasks are executed by registered workers. - Workflow output: a stable contract assembled from task output using
outputParameters. - Retry: task-scoped recovery. Retrying a failed
DO_WHILErestarts that loop's iteration history. - Pause and approval:
WAITandHUMANhold durable execution state until they are resolved. - Dynamic task / fan-out:
DYNAMICselects a task at runtime;FORK_JOIN_DYNAMICcreates runtime branches and is followed byJOIN. - Loop retention:
keepLastNbounds storage for long loops by intentionally removing older iteration history.
Verify before advising
Treat source as the specification. Check Java task and API implementations for runtime semantics, then check the relevant SDK source for SDK-authored agent, guardrail, and eval behavior. Run a JSON syntax check, a strict docs build, link validation, and a local execution whenever the configured server and integrations are available.
For machine-readable discovery, start at llms.txt. The curated llms-full.txt is generated from the source pages listed in the repository manifest.