Copewell: A Multi-Agent Swarm Architecture for Equitable Mental Wellness Support¶

🕒 Published (v1): 2026-07-02 14:40 UTC · Source: Arxiv · link

Why this paper was selected

Multi-agent swarm architecture for mental wellness; practical multi-agent system with equity framing

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

Copewell is a multi-agent swarm system for scalable mental wellness support that routes users to specialized AI agents based on continuous valence-arousal emotion mapping derived from Russell's Circumplex Model. It integrates self-reported, physiological, and contextual data streams to reduce assessment bias, and embeds ethical oversight as a first-class architectural agent rather than a post-hoc filter. The paper is a design and sociotechnical analysis informed by early practitioner engagement, not an empirical evaluation.

Problem¶

Current digital mental health tools overwhelmingly rely on single-mode conversational interfaces that apply a uniform support approach regardless of a user's dynamic emotional state, yielding high abandonment rates and little measurable relief. Algorithmic bias from single-modality assessment (self-report only) risks misclassifying user need. Ethical safeguards are typically post-hoc content filters rather than architectural features. Simultaneously, 75% of individuals in low- and middle-income countries (LMICs) receive no mental health treatment due to workforce shortages, cost, and stigma—a structural access gap no current AI tool is architecturally designed to address equitably.

Method¶

Copewell implements a three-stage pipeline—assessment → routing → intervention—coordinated across a swarm of specialized LLM agents.

Multi-source assessment fuses three data streams with fixed weights: self-reported mood/linguistic features (40%), physiological/biometric signals from device APIs such as HRV and sleep quality (35%), and environmental/calendar contextual data (25%). Weights are derived from Ecological Momentary Assessment principles.

Valence-arousal mapping projects the fused signal onto Russell's two-dimensional Circumplex Model (valence × arousal), enabling calculation of emotional distances, session centroids, and variability scores as continuous quantities.

Quadrant-based agent routing assigns users to one of four specialized agents: - Facilitator (Q1: high arousal, positive valence) — goal-setting, structured activity - Stabilising (Q2: high arousal, negative valence) — de-escalation, breathwork, grounding - Motivator (Q3: low arousal, negative valence) — behavioral activation, CBT-informed gentle engagement - Reflector (Q4: low arousal, positive valence) — journaling, mindfulness, meaning-making

Real-time linguistic analysis monitors mid-conversation state shifts and triggers agent handoff with shared context retention for narrative continuity.

Ethics Supervisor agent is a dedicated swarm participant (not a filter) that observes all agent outputs in real time, enforces safety protocols, detects crisis signals, and can override routing or intervene in conversations. The authors draw an analogy to prefrontal regulatory networks in human cognition.

Dual-mode intervention combines conversational support with sensory wellness protocols (e.g., visual grounding content, guided breathwork) triggered primarily by the Stabilising agent.

Key Contributions¶

Multi-source assessment framework with theoretically motivated 40/35/25 weighting schema to mitigate single-modality algorithmic bias
Valence-arousal routing mechanism integrating affective computing theory (Russell's Circumplex Model) with multi-agent coordination in place of keyword or intent classification
Ethics Supervisor as a first-class architectural agent enabling real-time ethical oversight (ethics-by-design rather than post-hoc filtering)
Dual-mode (conversational + sensory protocol) intervention delivery differentiated by emotional quadrant
Participatory design process engaging mental health practitioners as co-designers, informing architecture decisions ahead of empirical evaluation

Results¶

This paper presents no quantitative benchmark results. Empirical claims are limited to: - Background figures cited from literature: depression symptom reduction effect sizes ~0.28–0.38, generalized anxiety ~0.26–0.30 for digital mental health apps generally (not Copewell) - "Early practitioner engagement and beta deployment inform design decisions" — no outcome metrics reported - Future empirical evaluation is explicitly flagged as a direction, not a completed step

Limitations¶

No controlled empirical evaluation of Copewell itself; all quantitative figures are drawn from prior literature on other tools
The 40/35/25 assessment weighting is described as an original design decision informed by EMA principles but is not empirically validated against alternative weightings
Physiological and calendar data integration depends on device API availability and user consent, likely limiting coverage in low-resource populations the system aims to serve
The paper does not report crisis escalation performance, false-positive/negative rates for the Ethics Supervisor, or abandonment metrics from beta deployment
Cultural and linguistic adaptability of the valence-arousal mapping across LMIC populations is asserted but not demonstrated
Agent handoff coherence (shared context retention) is described architecturally but not evaluated for user-perceived continuity

Relevance to Agentic AI / LLM Agents¶

Copewell is a concrete instantiation of the specialized-agent-swarm pattern—routing to domain-specific LLM agents based on a continuous affective state signal rather than intent classification—directly relevant to researchers studying multi-agent coordination and task routing in LLM systems. The Ethics Supervisor represents an architectural blueprint for embedding normative oversight as an active swarm participant, a design pattern applicable beyond mental health to any high-stakes agentic deployment. The paper also surfaces a persistent challenge for agentic AI in sensitive domains: how to maintain narrative coherence and user trust across agent handoffs, a problem that applies broadly to multi-agent conversational pipelines. It contributes to the broader line of work on equitable and responsible LLM agent deployment, particularly for applications targeting underserved populations.