A lot of agent integrations fail because they expose tools at the wrong grain size. The agent gets a drawer full of tiny actions: list records, fetch record, update field, create note, send message, add tag, close task. Technically, it can do the job. Practically, it has to chatter.
Tool chatter is expensive and fragile. Each call creates another place to misunderstand state, lose context, hit a limit, or need review. It also pushes product knowledge into the agent prompt. The agent has to know the recipe for the job instead of the product offering the job as a command.
Compound commands are better.
A compound command packages a common workflow into one intentional operation: “prepare renewal risk brief,” “triage unread support escalations,” “draft account follow-up from latest activity,” “archive stale leads with dry run,” “sync content pack to publishing queue.” It can still expose steps and options, but the product owns the workflow contract.
This is not the same as hiding everything in a black box. The command should return a plan, a diff, assumptions, skipped items, warnings, and audit output. The point is to put the product's domain knowledge where it belongs. If the vendor knows the safe sequence, the validation rules, and the edge cases, the agent should not have to rediscover them through ten generic calls.
Good compound commands improve both speed and trust. They reduce token use. They make approvals easier because the human can review the intended unit of work. They give product teams a place to encode best practice. They make logs more meaningful because the action has a business name instead of a chain of tiny mutations.
The design challenge is choosing the right command size. Too small, and the agent drowns in mechanics. Too large, and the command becomes opaque automation. The sweet spot is the unit of work a capable operator would naturally delegate.
“Pull these records” is usually too small. “Run the whole department” is fantasy. “Find the risky renewals, explain why, draft next actions, and wait for approval before sending anything” is a real delegated job.
Agent-native products will compete on these units of delegation. The API surface will matter less than the work surface.
This is part 5 of 10 in Agent-Native Tools.
