Install, authenticate, then print the agent skill
A local agent should not need private dashboard context to understand how to send paperwork. The CLI should install, authenticate with an email code, and print a compact skill that tells the agent what it may and may not do.
That skill should say the quiet part clearly: inspect templates first, run dry runs, wait for human approval, send only approved packets, and report failures immediately.
Read before sending
A good playbook asks the agent to list templates, read the chosen template, and show the dry-run output before sending. This keeps the human in the approval loop without making the human click through a dashboard for every mechanical step.
For custom packets, the same pattern applies. Save the draft locally, capture feedback, revise, read the final version, and only then prepare the send.
Track from the terminal
After sending, the agent should be able to check status, view audit events, remind the signer, cancel the agreement, and download the signed PDF. These commands let the agent complete the workflow without pretending the dashboard is the product.
The dashboard can still be useful for humans. The CLI is the operational surface for the agent.
Capture feedback at the failure point
When install, login, or sending breaks, the best time to capture feedback is while the failed command is still on screen. A CLI feedback command can store the exact command, expected result, actual result, and severity.
That turns agent failure into product telemetry instead of a vague Slack message. It also gives the next agent enough context to fix the workflow.