Give AI one job and limits you can review.

Declare the agent's context, tools, and review points inside the same system as the Apps, data, and deterministic workflows around it.

Use agents for judgment. Use workflows for rules.

Give the agent the context and actions for one job. Keep predictable steps explicit in the workflow. Stop for a person before a consequential action when the business requires it.

The source names what the agent can use.

The project states the agent's job, what it may read, which tools it may call, and where a person must review the result. The runtime executes those declared boundaries.

Example: An account agent can read an approved brief and draft the next step. A workflow can stop for review before it changes a record or sends a message.

Swirls owns the production work behind it.

Hosted execution, supported access enforcement, declared tool delivery, review pauses, and the run record.

Declare context

Supply the business information needed for this job.

Declare tools

Expose only the actions the job needs.

Stop for review

Route sensitive or uncertain output to a person before action.

A robot agent studies business records, presents a recommendation to a human reviewer, and sends only the approved result to a workflow
  1. Business context
  2. Declared tools
  3. Conversation
  4. Human review
  5. Agent run history
Read the technical docs

You own

The agent's job, its context, the actions it may take, the acceptance criteria, and every required review point.

Give the agent a job you can define.

Name the context, allowed actions, expected result, and the point where a person must decide.