PromenautThe orchestration patterns
Most multi-agent work is built from a handful of patterns, often combined:
Sequential
Agents run in a chain, each handing the next the output it declared. Good for work with clear stages, such as extract, analyse, draft.
Parallel, or fan-out
Many agents work at once over a list of items, and the results are gathered. Good for breadth, such as scoring each requirement in a set.
Hierarchical
A lead agent plans and hands work to specialists, then composes their results. Good for open-ended tasks that need breaking down.
Conditional
The path depends on what earlier steps found, so the work is routed as it goes rather than following a fixed script.
Human in the loop
The work pauses to ask a person to approve, choose or answer, then carries on with the decision.
Underneath them sit three things that separate a demonstration from production: durable state, so a run survives a restart; recovery, so a stalled or failed step is picked up again rather than lost; and observability, so you can see what each agent did and in what order.
What good orchestration needs, and how Promenaut does it
Coordinating agents reliably is a real engineering problem. A production orchestration layer checkpoints as it goes, recovers when infrastructure fails, lets a person pause, resume or stop a run, and records what happened in order.
Promenaut composes agents into teams and teams into workflows with declared inputs and outputs, fans work out over a task list an earlier step produced, checkpoints long-running work so a restart resumes rather than starts again, re-queues a workflow that has stalled, runs on a schedule, and lets a person pause, stop or cancel a run.
When a workflow is not confident enough to go on, it can ask a person and use the answer.
Why orchestration alone is not enough in a regulated enterprise
Orchestration answers how agents work together. On its own it does not answer what a regulated firm’s risk function and regulator will ask: who each agent acted as, and whether it could reach more than that person may; what each agent did, for any run, including the ones that failed; what a run will cost before it starts; and how to stop work cleanly.
This is why a framework or an orchestration library gets a team to a working multi-agent prototype, and then the programme stalls before production: the coordination is there, but the authority, evidence, cost control and containment are left for you to build.
Orchestration coordinates the work. It does not govern it. In a regulated setting, you need both.
From orchestration to a governed workforce
What a regulated enterprise needs is orchestration plus control, produced by the work itself. In Promenaut, the workflow that coordinates a team of agents also governs it.
In the same run, an agent reaches knowledge as a deputy of the person who launched it; sensitive steps wait for a person, and a sensitive knowledge change waits for a checker who did not launch it; the run is costed before it starts; each agent run records which model answered; each workflow run ends with a verdict in a tamper-evident ledger; and an operator can pause an agent or stop a running workflow.
Common questions
What is AI agent orchestration?
The coordination of several AI agents to complete work that is too much for a single agent alone: sequencing them, passing work between them, handling branching, shared state and failure, and deciding where a person steps in.
What are the main agent orchestration patterns?
Sequential, parallel or fan-out, hierarchical, conditional, and human in the loop, usually combined, over durable state with recovery from failure.
How is orchestration different from a single AI agent?
A single agent plans and calls tools for one task. Orchestration composes several agents into a team that carries a whole piece of work through several steps to a finished result.
Is orchestration enough to run agents in a regulated enterprise?
No. Orchestration coordinates the work but does not govern it. A regulated firm also needs authority per run, evidence, cost control and a clean way to stop work, which is what turns orchestration into a governed digital workforce.
