Where do the first agents come from?
After a few months of experimenting, many companies have agents in support, sales, finance and operations, each built where one team saw a clear pain. That is a reasonable way to start.
Why do they overlap?
Most work moves through several teams. Customer feedback is triaged in support, assessed by product, picked up by engineering and only then turned into a reply. Onboarding a vendor touches procurement, legal, finance, security and IT.
When several teams look at the same flow from their own side, they can each build an agent for the same step: on different stacks, with different prompts, owned by people who don't know about each other. Sometimes two of them change the same record and get in each other's way.
What does agent orchestration need?
For one agent to hand work to the next, someone has to define the next step, who owns it, and what arrives in what shape. That is a description of the process, not a feature of a tool.
A map of the end-to-end flow, with owners, handoffs and the steps that repeat across flows, often shows that several agents are really one shared step. That step can then be designed once and reused wherever it appears.
This is the map kawaru.ai builds: how the company works across the whole flow, and where AI could fit next.