Agents
Orchestration
Someone has to set the tempo when several steps, tools, or agents work together: sequence, handoffs, error handling, restart. Orchestration is the unglamorous layer where demos become systems. Whoever builds it cleanly can swap out models without reinventing the workflow.
An example
A complaint passes through four stations: checking eligibility, calculating the credit, drafting the reply, approval. The orchestration is what stands in between: it calls the stations in the right order, passes results along, catches errors, retries on timeout, aborts after three failed attempts, and logs every step. No model does this. It is ordinary, boring, indispensable software.
A closer look
Orchestration determines which step runs when, what happens on errors, where a human is asked, and how intermediate states are stored. In practice this is the larger part of the work. The model is only one building block in it.
Why it matters
In projects that want to get from pilot to operation, this is where the biggest underestimated effort sits. The pilot works because a human sits next to it and steps in when something sticks. In operation nobody sits next to it, so every failure case has to be decided beforehand: what happens on a timeout, on duplicate processing, on a system outage in the middle of the workflow? These questions are familiar from thirty years of process automation and are regularly overlooked anew in AI projects.
Commonly misunderstood
Many initiatives fail not because of the model but because of missing resumption: a workflow breaks off at step seven, and nobody knows whether step six has already sent an email.