aris/docs/ai-agent-ideas.md

AI Agent Ideas for ARIS

Brainstorm document. Not all ideas are feasible or desirable - just capturing possibilities.

1. Feed Curation Agent

Sits between the reconciler and UI. Reranks/filters the raw feed based on learned preferences and context.

Examples:

• User always dismisses weather items in the morning → agent deprioritizes them
• User frequently taps calendar items before meetings → agent boosts them 30 minutes prior
• Deduplicates or groups related items ("3 meetings in the next hour")
• Learns time-of-day patterns (work items in morning, personal in evening)

Interface:

interface FeedAgent {
	process(
		items: FeedItem[],
		context: Context,
		userPreferences: UserPreferences,
	): Promise<FeedItem[]>
}

Fits naturally as a post-processor after reconciliation.

---

2. Query Agent

User asks natural language questions about their feed or connected data.

Examples:

• "What's on my calendar tomorrow?"
• "When's my next flight?"
• "Summarize my day"
• "Do I have any conflicts this week?"
• "What did I have scheduled last Tuesday?"

Behavior:

• Agent queries relevant sources directly or searches recent feed history
• Synthesizes data into conversational response
• Could generate a temporary filtered feed view

---

3. Proactive Agent

Monitors context changes and triggers actions without user prompting.

Examples:

• Near grocery store + "buy milk" in tasks → surfaces reminder
• Calendar conflict detected → alerts before it happens
• Weather changing → suggests leaving earlier for commute
• Unusual traffic on commute route → notifies user
• Meeting in 10 minutes but user hasn't moved → gentle nudge
• Package delivered + user is home → notification

Implementation considerations:

• Needs background processing / push capability
• Privacy implications of continuous monitoring
• Battery/resource usage on mobile

---

4. Source Configuration Agent

Helps users set up and tune sources through conversation.

Examples:

• "Show me fewer emails"
• "Only show calendar events for work"
• "I don't care about weather"
• "Prioritize tasks over calendar"
• "Add my Spotify account"

Behavior:

• Translates natural language into source config changes
• Can explain what each source does
• Helps troubleshoot when sources aren't working

---

5. Feed Item Generation Agent (AI-Native Sources)

Some sources are AI-powered rather than API-driven.

Examples:

• Daily briefing: "You have 4 meetings today, busiest is 2-4pm"
• Pattern-based reminders: "You usually go to the gym on Tuesdays"
• Suggested actions: "You haven't responded to Sarah's email from yesterday"
• Weekly review: "You completed 12 tasks this week, 3 are overdue"
• Context synthesis: "Your flight lands at 3pm, you have a meeting at 5pm - that's tight"

These are sources that implement DataSource but use an LLM internally.

---

6. Action Agent

Executes actions on behalf of the user based on feed items.

Examples:

• "Snooze this reminder for 1 hour"
• "RSVP yes to this event"
• "Mark this task as done"
• "Send a quick reply saying I'll be late"
• "Book an Uber to this location"

Considerations:

• Needs action capabilities per source
• Confirmation UX for destructive/costly actions
• OAuth scopes for write access

---

7. Explanation Agent

Explains why items appear in the feed.

Examples:

• User asks "Why am I seeing this?"
• Agent explains: "This calendar event starts in 15 minutes and you marked it as important"
• Helps users understand and trust the system
• Useful for debugging source behavior

---

8. Cross-Source Reasoning Agent

Connects information across multiple sources to surface insights.

Examples:

• Calendar shows dinner reservation + weather source shows rain → "Bring an umbrella to dinner"
• Flight delayed + calendar has meeting after landing → "Your 3pm meeting may be affected by flight delay"
• Task "buy birthday gift" + calendar shows birthday party tomorrow → boosts task priority
• Email mentions address + maps knows traffic → "Leave by 2pm to make your 3pm appointment"

This is more complex - requires understanding relationships between items.

---

9. Memory Agent

Maintains long-term memory of user interactions and preferences.

Examples:

• Remembers user dismissed a recurring item 5 times → stops showing it
• Knows user's home/work locations from patterns
• Tracks what times user typically checks the feed
• Remembers user's stated preferences from conversations
• Builds implicit preference model over time

Feeds into other agents (especially Feed Curation).

---

10. Onboarding Agent

Guides new users through setup conversationally.

Examples:

• "What apps do you use for calendar?"
• "Would you like to see weather in your feed?"
• "What's most important to you - tasks, calendar, or communications?"
• Progressively enables sources based on conversation
• Explains privacy implications of each source

---

11. Anomaly Detection Agent

Surfaces unusual patterns or items that break routine.

Examples:

• "You have a meeting at 6am tomorrow - that's unusual for you"
• "This is your first free afternoon in 2 weeks"
• "You haven't completed any tasks in 3 days"
• "Your calendar is empty tomorrow - did you mean to block time?"

---

12. Delegation Agent

Handles tasks the user delegates via natural language.

Examples:

• "Remind me about this tomorrow"
• "Schedule a meeting with John next week"
• "Add milk to my shopping list"
• "Find a time that works for both me and Sarah"

Requires write access to various sources.

---

13. Summary Agent

Generates periodic summaries of feed activity.

Examples:

• Morning briefing: "Here's your day ahead"
• Evening recap: "Here's what happened today"
• Weekly digest: "This week you had 12 meetings, completed 8 tasks"
• Travel summary: "Your trip to NYC: 3 flights, 2 hotels, 5 meetings"

Could be a scheduled AI-native source.

---

14. Notification Agent

Decides what deserves a push notification vs. passive feed presence.

Examples:

• High-priority items get pushed
• Learns what user actually responds to
• Batches low-priority items into digest notifications
• Respects focus modes / do-not-disturb

Reduces notification fatigue while ensuring important items aren't missed.

---

15. Conversation Agent

General-purpose assistant that can discuss feed items.

Examples:

• User taps an item and asks "Tell me more about this"
• "What should I prepare for this meeting?"
• "What's the best route to this location?"
• "Who else is attending this event?"

Contextual conversation anchored to specific feed items.

---

Implementation Priority Suggestion

If implementing incrementally:

1. Feed Curation Agent - highest value, fits existing architecture
2. Query Agent - natural user expectation for AI assistant
3. Summary Agent - low risk, high perceived value
4. Proactive Agent - differentiator, but complex
5. Cross-Source Reasoning - advanced, builds on others

---

Open Questions

• Where do agents run? (Client, server, edge?)
• How to handle agent latency in feed rendering?
• Privacy model for agent memory/learning?
• How do agents interact with third-party sources?
• Cost management for LLM calls?