SENSE
SENSE is a consent-first agentic system for crisis prevention.
Core trick
Guardian agent checks every outbound message against the user's pre-authorized consent scope and cadence limits, blocking violations and logging each decision.
Description
SENSE is a consent-first agentic system for crisis prevention. During a stable moment, a user authors their own Stanley-Brown safety plan and approves exactly what the system may do during predictable high-risk windows like birthdays. Agents then execute that plan: a Companion Agent sends one evidence-based Caring Contacts message (brief, warm, non-demanding, the only messaging intervention with RCT support), while a Guardian agent validates every outbound action against the user's consent scope and cadence limits, blocking violations and writing an append-only audit trail. Not a chatbot, there is no chat. SENSE is an agent system whose headline feature is what it provably refuses to do. Built with the Anthropic API. Future work: Forecast, Circle, and Bridge agents per the repo README.
README excerpt
# SENSE Anthropic x Abridge x Lightspeed Hackathon, July 2026 A consent-first crisis prevention agent system. People protect their future selves during predictable high-risk windows (birthdays, anniversaries, holidays) by building their own safety plan during a calm, stable moment — agents only ever act on what was explicitly authorized then. Full evidence citations (Caring Contacts RCT, Stanley-Brown Safety Planning) and the complete build order land in this README as each phase ships. See commit history for progress. ## Design principles 1. **Consent-first, always.** No action is taken on inferred mental state. Every automated action traces back to a `Consent` the user granted during onboarding, and can be revoked with one tap. 2. **Caring Contacts, not message floods.** Brief, warm, non-demanding, infrequent outreach (Motto & Bostrom). At most 1 message/day/channel, 2-3 total per risk window. 3. **Human-in-the-loop escalation.** Agents draft, schedule, and recommend. A human (the user or their chosen contact/clinician) approves high-stakes actions. 4. **Stanley-Brown Safety Planning as the data model.** Agents operationalize the user's own plan; they do not invent interventions. 5. **Privacy by architecture.** Local SQLite, field-level encryption at rest, no third-party ad-tech, one-tap deletion. 6. **Crisis fallback is always visible.** 988 and Crisis Text Line (text HOME to 741741) appear in every view and every outreach message. ## Setup