Vision
Move Luna from reactive (responds when messaged) to proactive (surfaces things unprompted). The Jarvis model: Luna monitors feeds and signals continuously, decides what's worth attention, and messages the owner without being asked.
Problem with the current model
Right now Luna can only act when:
- A user sends a message, or
- A scheduled reminder fires
Both are pull-based. True proactive behavior requires push-based loops: "something changed → evaluate → notify if worth it."
Proposed: always-on monitoring agents
Introduce lightweight polling agents that run on tight cron schedules and evaluate signals against a relevance threshold before notifying.
Example monitors:
- Telegram importance filter — scan new inbound messages while owner is away; surface only the ones that are urgent or from high-priority contacts. "Sardor sent 3 messages in the last hour, looks like something's wrong."
- Calendar conflict watcher — poll Google Calendar every 15 min; alert if a meeting was rescheduled, added, or conflicts with another.
- Email triage — scan inbox every N minutes; surface emails from VIPs or containing keywords like "urgent", "invoice", "deadline".
- Home inventory / grocery — integrate with a shopping list or smart home sensor; remind to restock when something runs low.
- News/market alert — watch for breaking news on configured topics (Uzbekistan fintech, AI, portfolio companies); surface only high-signal items.
Architecture
Monitoring loop (cron, tight interval e.g. */15 * * * *)
→ Poll data source (Telegram API, Gmail, Calendar, etc.)
→ Filter: is this worth the owner's attention?
Yes → send_message to owner DM with context
No → discard silently
→ Update "last checked" state in memory
Key properties:
- Stateful — each loop tracks last-seen state (last message_id, last calendar event, etc.) in memory to avoid re-alerting on the same item
- Threshold-based — not every new email is worth a ping; the agent evaluates priority before sending
- Low-noise — better to miss one alert than to spam the owner; false positives destroy trust faster than false negatives
- Composable — each monitor is an independent scheduled subagent, not a monolithic loop
Trust / safety considerations
Proactive agents acting on real data (sending messages on the owner's behalf, booking things, managing calendar) need tighter permission gates than reactive responses. Recommend:
- Explicit owner-approval step for any write action (send, book, cancel)
- Read-only monitors by default; destructive actions require a confirmation message first
- Audit log of every proactive ping sent, so the owner can tune the threshold
Related
Vision
Move Luna from reactive (responds when messaged) to proactive (surfaces things unprompted). The Jarvis model: Luna monitors feeds and signals continuously, decides what's worth attention, and messages the owner without being asked.
Problem with the current model
Right now Luna can only act when:
Both are pull-based. True proactive behavior requires push-based loops: "something changed → evaluate → notify if worth it."
Proposed: always-on monitoring agents
Introduce lightweight polling agents that run on tight cron schedules and evaluate signals against a relevance threshold before notifying.
Example monitors:
Architecture
Key properties:
Trust / safety considerations
Proactive agents acting on real data (sending messages on the owner's behalf, booking things, managing calendar) need tighter permission gates than reactive responses. Recommend:
Related