September 2026 Summaries

Conductor’s AGENT task integrates AI agents into durable workflows, allowing them to operate alongside HTTP calls, human approvals, parallel branches, and other task types while gaining workflow features such as retries, timeouts, state persistence, observability, and error handling. It supports external A2A-compatible agents through the default `a2a` mode, which can use a direct endpoint or agent-card URL, and agents deployed on Conductor through `conductor` mode, including agents built with supported SDKs or imported from frameworks such as OpenAI Agents, LangGraph, and Google ADK. When an agent does not return immediately, Conductor records its remote task identifier and repeatedly polls for completion; completed results provide response text, state, task and context identifiers, and, for Conductor-hosted agents, a child-workflow execution ID. These identifiers enable multi-turn conversations and human-in-the-loop flows when an agent requests more input, while workflow branching can respond to states such as completed, failed, canceled, or input-required. Configurable duration and polling-failure limits prevent workflows from waiting indefinitely and let standard workflow error handling manage unavailable or slow agents.

Conductor positions its platform as a unified durable execution engine for both AI agents and workflows, allowing agents to operate as workflow components with persistent memory, retries, approval handling, exact-once execution, and long-term execution histories. It argues that production agentic systems require deterministic controls around nondeterministic models, with workflows predefining allowed tools, branches, validation, and actions while models remain responsible for judgment within those constraints. The platform supports agents built with frameworks including LangChain, LangGraph, OpenAI, Vercel, and Google ADK, as well as agents written through Conductor SDKs, and enables tools to include APIs, functions, MCP tools, human approvals, other agents, or complete workflows. Conductor emphasizes determinism before execution, durability during execution through engine-managed state that survives failures and delays, and explainability afterward through unified records of prompts, model responses, tool calls, inputs, outputs, and approvals. Its approach treats agents and workflows as interchangeable parts of a single system, enabling workflows to invoke agents for judgment-based tasks and agents to invoke workflows for reliable, repeatable actions.