Start with channel cost.
A text message should require stronger evidence than a digest entry or review queue item.
How to score evidence before an AI agent texts you
A proactive text is one of the most expensive actions a personal AI agent can take. Before it sends, make it score the evidence, explain the threshold, and preserve a correction path.
source qualitydeadline proximitysender authorityfreshnessreversibilityprior correctionsource qualitydeadline proximitysender authorityfreshnessreversibilityprior correction
Score the evidence, not just the model output.
The agent should know when a confident summary still does not deserve a text. Evidence scoring makes that decision inspectable.
Use a readable receipt.
Show the top evidence, the threshold, and the correction path in language the user can judge.
Text lane
Use with a text message AI assistant when the score clears the highest threshold.
Browser lane
Send qualified but non-urgent work into computer-use cache workflows.
Build lane
Require stronger receipts when an AI agent builds websites or edits public pages.
Collect the candidate signal.
Capture the message, source, timestamp, sender, related calendar entry, prior receipt, and any user correction that applies.
Score each factor.
Rank source quality, freshness, deadline proximity, sender authority, reversibility, and known user tolerance from low to high.
Compare against channel thresholds.
Text requires the highest score. Review queue requires medium score. Digest accepts lower scores. Ignore stale or weak signals.
Attach the receipt.
Every proactive text should include or link to a receipt explaining why the agent chose SMS instead of batching.
Train on corrections.
When the user marks a text as unnecessary, early, late, duplicate, or missed, update the system prompt with what went wrong and what should happen instead.
Evidence scorecard template
| Factor | Question | Text threshold |
|---|---|---|
| Source quality | Is the signal first-party, direct, and fresh? | High |
| Urgency | Will waiting materially harm the user? | High |
| Reversibility | Can the action be undone if the agent is wrong? | Medium to high |
| User preference | Has the user asked for this category to interrupt? | High |
| Correction history | Has this pattern caused false positives before? | Must be reviewed |
Implementation checklist
- Separate evidence score from model confidence.
- Set a higher threshold for text than for digest or review.
- Require the agent to explain why the item matters now.
- Log source, timestamp, factor weights, final route, and receipt ID.
- Offer one-tap correction labels: useful, too early, too late, duplicate, unnecessary, missed.
- Review false positives weekly before expanding autonomy.
Prompt clause
Before sending a proactive text, score the evidence by source quality, urgency, freshness, reversibility, user preference, and correction history. If the evidence does not clear the SMS threshold, route the item to review or digest. Always provide a receipt and correction path.
Sources
- Nielsen Norman Group on notification design for alert clarity and user-control principles.
- Microsoft Research on interrupted work for attention-switching costs.
- NIST AI Risk Management Framework for transparent AI risk-management framing.
- Supers for personal AI agent workflows across text, browser work, and generated outputs.
FAQ
What score should trigger a text?
Use a high threshold and tune it from user corrections. SMS should be reserved for strong, fresh, user-relevant evidence.
What happens below the threshold?
Route to review, digest, or ignore depending on the score and user preference.
Should the user see the score?
Show a readable receipt with the top reasons. The exact numeric score can stay internal unless the user asks.
Where should this be tested first?
Start with a text message AI assistant, then reuse the same receipts for browser and page-building workflows.
Make every AI text earn its interruption.
Supers-style personal agents become more trustworthy when proactive messages carry evidence receipts, correction labels, and channel thresholds.