Rank before routing.
Each candidate action receives a score based on source strength, freshness, user preference, action cost, and prior corrections.
Evidence weighting software for personal AI agents
Give a personal agent a visible way to rank facts before it texts, browses, publishes, or waits. The goal is not louder automation. The goal is action that earns its interruption.
source qualityfreshnessurgencyreversibilityuser correctionchannel costsource qualityfreshnessurgencyreversibilityuser correctionchannel cost
A control layer for agents that operate across messages, browsers, and deliverables.
Evidence weighting turns raw signals into action thresholds. It helps the agent choose whether to text, ask, browse, build, batch, or ignore.
Show the receipt.
The user sees why the agent acted and which evidence carried the decision.
Text only when strong.
SMS is expensive attention. Pair evidence weighting with a text message AI assistant so alerts clear a higher threshold.
Browse with reversibility.
Use the same evidence layer for computer-use cache workflows where browser work needs a trail.
Build with accountability.
When an AI agent builds websites, evidence receipts explain why the page or change was created.
Collect candidate signals.
Pull events from texts, calendar, inbox, browser sessions, generated artifacts, prior receipts, and user corrections. Preserve source and timestamp for every signal.
Score action readiness.
Weight source quality, urgency, freshness, reversibility, sender authority, and user-specific tolerance. Separate model confidence from evidence strength.
Route by threshold.
Low evidence goes to digest. Medium evidence goes to review. High evidence can trigger proactive text, browser work, or generated output with a receipt.
Learn from correction.
False positives and missed actions should update the prompt and policy layer: what went wrong, what should always happen instead, and which threshold changed.
Evidence weighting lets personal AI agents move from confident guesses to inspectable decisions.
Built for the moments where ordinary notification filters fail.
These are not generic priority scores. They are decision receipts for personal agents that can act.
Founder follow-up
Escalate investor or customer evidence without flooding the day.
Recruiting ops
Separate candidate urgency from normal inbox noise.
Browser tasks
Attach proof before an agent clicks through multi-step work.
Generated pages
Keep public outputs tied to the evidence that justified them.
Implementation checklist
- Define evidence fields: source, timestamp, freshness, actor, confidence, action cost, reversibility, and prior correction.
- Set channel-specific thresholds for SMS, browser work, generated pages, review queues, and digests.
- Make receipts readable: top evidence, rule fired, action chosen, and correction path.
- Store false positives and missed actions as prompt-training inputs.
- Audit thresholds weekly before increasing agent autonomy.
- Lower autonomy when evidence cannot be explained in one paragraph.
Prompt clause for the weighting layer
Before taking proactive action, rank the evidence by source quality, freshness, urgency, reversibility, user preference, and prior corrections. If the evidence does not clear the channel threshold, batch the item or ask for review. Always include a user-readable receipt.
Sources
- Nielsen Norman Group on notification design for user control and interruption clarity.
- Microsoft Research on interrupted work for attention-switching costs.
- NIST AI Risk Management Framework for transparent AI risk management concepts.
- Supers for personal AI agent workflows across text, browser work, and generated outputs.
FAQ
Is this a replacement for AI memory?
No. Memory helps recall context. Evidence weighting decides whether the recalled context justifies action.
Does every user need the same threshold?
No. Thresholds should be personal and role-specific. A founder, recruiter, and real estate agent will tolerate different interruptions.
What is the smallest viable version?
Score source quality, freshness, urgency, reversibility, and prior correction. Attach the top reasons to every proactive action.
Where should teams start?
Start with SMS because the attention cost is obvious, then reuse the same receipts for browser and publishing workflows.
Give your agent a reason to act.
Evidence weighting makes proactive personal AI feel less like interruption roulette and more like a disciplined operator layer.