what is orchestration in ai?

the three things the word covers

Model orchestration. Routing requests between providers and model sizes. Send the cheap model the easy request, fall back when one provider is down, cache repeated calls. This is a cost and reliability concern, and it is invisible when it works.

Pipeline orchestration. Sequencing the steps around a model call: fetch, chunk, embed, retrieve, rerank, generate, validate, store. Classic data engineering with an inference step in it. Airflow, dagster, temporal, and every rag framework live here.

Agent orchestration. Coordinating several autonomous agents toward a goal, where the steps are not fixed in advance. This is the newest sense and the one people usually mean now.

Half the confusion in any discussion about ai orchestration is two people using two of these three definitions.

what every orchestration layer has to answer

Regardless of which sense you mean, the same five questions:

1. What runs next, and who decides: a diagram or an agent.
2. What data does it get, and how much history rides along.
3. What happens on failure: retry, fall back, escalate, or stop.
4. When is it finished, including the case where it never converges.
5. What is recorded, so a wrong result can be explained afterwards.

A system missing an explicit answer to number three or number four will eventually cost you a surprising invoice.

fixed path or chosen path

The one distinction worth holding on to. Pipeline orchestration fixes the path in advance, so the same input always takes the same route: cheap, fast, and debuggable. Agent orchestration hands over a goal and lets the path be chosen at runtime: flexible, more expensive, and harder to audit.

Neither is better. Using an agent where a pipeline would do buys nondeterminism you did not need. Using a pipeline where the branches are long-tailed means enumerating cases forever. Most systems that survive are hybrid: a deterministic skeleton with agent judgement at the few nodes that need it.

Workflow versus agentic orchestration covers where that line falls in practice.

how this works in aldena

Aldena is agent orchestration with deterministic edges. The org chart decides who work goes to, and a manager decomposes and delegates down its own line rather than following a diagram you drew. Inside a task, agents choose their own steps.

The five questions above are answered by the product rather than by your code. Failure and finish conditions are explicit: irreversible actions stop at an approval gate that is on by default. Data flow between agents rides on shared room memory, 100 entries every agent reads plus 50 private each, so a handoff carries facts rather than a summary. And the record is the room itself, in an isolated server you can watch live.