November 2026 Summaries

Orkes Conductor has introduced an enterprise add-on for multi-region and multi-cloud resilience that automatically replicates workflow execution state and configuration to a secondary cluster, monitors cluster health, and redirects a shared endpoint during failures without requiring client, agent, worker, or tool reconfiguration. Designed to protect mission-critical agentic and conventional workflows from regional or cloud-provider outages, the capability can resume in-progress work on the replica cluster, though business-action tasks should be idempotent because some tasks may be retried during recovery. It extends existing worker-restart and multi-availability-zone protections by supporting two active deployment locations across regions or providers, with encrypted replication and data remaining in customer-selected regions, clouds, and accounts. Orkes targets a 10-minute recovery time objective, a maximum 90-second recovery point objective for asynchronous workflows, and near-zero data loss for synchronous workflows during failover. The feature is available for new and existing Orkes-hosted or customer-hosted clusters across AWS, Microsoft Azure, and Google Cloud.

A custom Conductor workflow is presented as a pre-install alternative to npm audit for evaluating public npm packages before adoption, addressing npm audit’s post-install timing, noisy transitive-dependency findings, and limited contextual guidance. The auditor gathers package metadata, dependency graphs, vulnerabilities, licenses, deprecation status, provenance attestations, repository health indicators, maintenance notes, and Hacker News reports from sources including npm, OSV.dev, deps.dev, DeepWiki, and Hacker News, then uses an LLM to summarize evidence while fixed routing rules determine whether to approve, reject, or send a package for human review. Its four nested workflows assess packages, scan dependency trees for factual security data, record decisions as GitHub issues, and coordinate the full adoption review; read-only scanning and assessment workflows are also exposed as MCP tools for agents and CI/CD systems. The design uses parallel API calls and sub-workflows to scale to large dependency trees, avoids MCP timeout issues for longer LLM assessments, and ensures that uncertain cases receive human review rather than relying solely on AI. Suggested extensions include incorporating additional community sources, considering intended package use, reviewing entire lockfiles, and exposing full approval decisions through MCP, while the broader pattern is proposed for other high-stakes approval processes.

Orkes Conductor can build turn-based AI chatbots by combining LLM-powered agents with HUMAN tasks that pause a workflow until a user responds, allowing an entire multi-turn conversation to persist in one durable execution without requiring separate worker processes. Developers define an agent with a model, instructions, and a single human_tool such as ask_human, which the agent must be instructed to call on each conversational turn; Conductor compiles the agent into a looping workflow that alternates between model calls and user-response pauses. An external chat client drives the interaction by polling for paused executions, displaying the agent’s message, collecting the user’s reply, and resuming the same execution. Conversations can end when a user exits, when the agent produces a final response without calling a tool, or when a configured maximum turn limit is reached. The hosted, free Orkes Conductor Developer Edition provides the necessary environment, requiring users to create an account, generate an access key, configure an LLM provider, deploy the agent, and start a conversation.