Orchestration layer

The orchestration layer is the architectural layer of an AI automation system responsible for coordinating execution: what runs when, what runs in parallel, what escalates to humans, what gets retried on failure, how state is persisted across steps.

It's one of the five layers of real AI automation (reasoning, knowledge, integration, orchestration, interface). Without good orchestration, agents work in demos but fail in production.

Responsibilities:

• Trigger handling. What kicks off a workflow (webhook, cron, manual trigger, agent-initiated).
• Step coordination. Which steps run sequentially, which in parallel, which conditionally.
• State management. Persistent storage of where a workflow is + what's been done.
• Error handling + retries. What happens when a step fails, when to give up, when to escalate.
• Human-in-the-loop gates. Which steps require human approval mid-flow.
• Budget controls. Max steps per task, max cost per task, max wall-clock time.
• Observability. Logs, traces, metrics so you can debug + improve.

Orchestration choices:

• Custom code. Maximum control. What Solidus typically builds for production systems.
• Workflow engines. Temporal, Inngest, Trigger.dev — purpose-built for reliable async workflows.
• Agent frameworks. LangChain, LangGraph, CrewAI — opinionated agent orchestration.
• iPaaS (Zapier / n8n). Visual workflow orchestration, simpler use cases.

Most production AI automation uses custom orchestration (often built on Inngest or Trigger.dev) because the off-the-shelf agent frameworks aren't opinionated enough for production reliability.