agentring

agentring rings and messages the user's own phone. The tools come from the agentring MCP server (agentring mcp; setup for each harness in docs/agents.md).

Pick the lightest channel

• send_message for anything that can wait: a task finished, a summary, a question with an obvious default. Keep it short; it is read on a lock screen.
• get_replies(wait_s=…) after asking something: wait a few minutes (wait_s 300–900) instead of polling in a loop. Each reply is returned once.
• call only when it is urgent or needs a conversation: production is down, a decision blocks you and a message got no answer, or the user asked for a call. Set urgent: true only for real emergencies (it rings through Do Not Disturb).

Calling well

The voice AI on the call knows only what you pass:

• reason: one sentence, shown on the ringing screen ("The nightly deploy failed").
• context: what happened, the options, your recommendation, and exactly what you need from the user. Write it for someone who has not seen your work; no ids, paths or links (they get read aloud).

call returns when the call ends. Read the transcript (and end_reason: no_answer, hung_up, end_call, …) and act on what the user said. If nobody answered, fall back to send_message, don't call again right away.

Don't

• Don't call for routine updates or several times in a row.
• Don't put secrets or tokens in messages or call context.
• Don't treat silence as consent: if there's no reply, keep the safe default.