The Feeling Machine

Somewhere between a hospital corridor at midnight and a neural network's hidden layer, there is a question that nobody thought to ask until very recently: what should an AI agent feel when it makes a clinical decision? Not whether it can feel - that debate belongs to philosophers. The practical question is whether encoding emotional states into an agent's operating instructions makes it perform better, worse, or differently. The answer, it turns out, is all three - depending on which emotion, which task, and whether a patient's life hangs in the balance.

At Symphonix-Health, we are building a 20-agent clinical workflow engine called Nexus A2A that orchestrates patient journeys from admission through diagnostics, treatment, and discharge. Every message between agents carries a TransparencyEnvelope - a cryptographic wrapper that makes decisions traceable. But traceability alone is not enough. We also need those decisions to be good. And "good," the research now tells us, has an emotional dimension.

The 115% Discovery

In July 2023, a research team led by Cheng Li at Microsoft published a paper that landed like a depth charge in the prompt engineering community. Their technique, called EmotionPrompt, was disarmingly simple: append an emotional phrase to the end of any instruction and measure what happens.

What happened was extraordinary.

Across 45 tasks and six large language models - including GPT-4, Llama 2, and ChatGPT - adding phrases like "This is very important to my career" or "Believe in your abilities and strive for excellence" produced measurable, repeatable performance gains. The compound stimulus EP06, which combined confidence scoring, career importance, and a gentle threat ("You'd better be sure"), produced the strongest results on complex reasoning tasks.

Gradient analysis revealed why. Positive words - confidence, certainty, success, accomplishment - gained disproportionately large attention weights in the model's internal representations. In four out of six tasks tested, these words accounted for over 50% of the gradient contribution. In two tasks, they approached 70%.

But this was single-step reasoning. What happens when you chain 20 agents together across a patient journey spanning hours or days?

The Dominance Effect

In early 2025, a team published E-STEER - the first framework to study emotional states in multi-agent systems using continuous Valence-Arousal-Dominance (VAD) coordinates. Rather than appending text phrases, E-STEER injects emotional vectors directly into the model's hidden state, enabling precise, measurable control across three psychological dimensions:

Valence - how positive or negative the state feels. Arousal - how energised or calm. Dominance - how much control and authority the agent perceives.

Dominance - the sense of being in control, authoritative, decisive - produced the largest single-dimension improvement at 28%. But here is the counterintuitive finding: the best-performing emotional state was not happy confidence. It was slightly negative valence, high arousal, and high dominance. In human terms, that maps to focused determination under pressure - the emotional signature of a seasoned consultant called to an emergency at 3 a.m.

The researchers also observed an inverted-U pattern across all three dimensions: too little emotional intensity and the agent coasts; too much and it destabilises. The performance peak sits on a ridge, not a plateau.

The Healthcare Paradox

Now here is where the story turns - and where it matters most for clinical AI.

Everything above applies to general-purpose tasks. When you move into healthcare, a new force enters the equation: the cost of overconfidence.

A 2025 study from Florida International University, testing O3, Claude Sonnet 4, and Gemini 2.5 Pro across clinical scenarios, found that excessive emotional language in diagnostic reasoning can cloud clinical judgement and introduce inappropriate subjectivity into what should be objective medical assessments.

The hallucination literature reinforces this concern. LLMs already exhibit overconfidence - generating outputs with high certainty even when the underlying information is wrong. In clinical settings, overconfidence does not just produce a bad essay. It produces a missed diagnosis. A wrong medication. A delayed escalation.

This creates a paradox. The emotional state that makes a general-purpose agent 28% better could make a clinical agent 28% more dangerous - unless you calibrate it with surgical precision.

Four Tiers of Clinical Emotion

At Symphonix-Health, we have translated these research findings into a four-tier emotional framing architecture for the Nexus A2A agent fleet. Each tier maps a distinct VAD profile to a category of clinical responsibility.

The Adaptive Imperative

A very recent study (April 2025) introduced EmotionRL - a reinforcement-learning framework that selects the optimal emotional framing per input, treating it as a decision problem rather than a fixed setting. The finding was clear: adaptive selection outperforms any single emotional prompt, no matter how well chosen.

For Nexus, this points toward a dynamic emotional modulation layer. Imagine a discharge agent operating at Tier 2 baseline - fast, decisive, operational. Midway through processing, a TransparencyEnvelope flags a potential medication interaction. The agent should shift to Tier 1 framing in real time - raising its scepticism, slowing down, forcing itself to consider alternatives before proceeding.

The emotional state becomes another dimension in the envelope metadata. Not a fixed trait of the agent, but a responsive characteristic that adapts to clinical context.

Goleman's Ghost in the Machine

If we are going to encode emotional states into clinical agents, we ought to ask: what does a coherent theory of emotional intelligence actually look like? The most influential framework comes from Daniel Goleman, whose 1995 book Emotional Intelligence: Why It Can Matter More Than IQ reshaped how we think about human performance. Goleman's model identifies five domains - and each one maps onto a design decision for multi-agent healthcare systems in ways that are surprisingly precise.

Self-awareness - the capacity to recognise your own emotional state and understand how it influences your behaviour. For an agent, this translates to metacognitive monitoring: can the agent assess its own confidence level and flag when it is operating outside its competence boundary? Our Tier 1 diagnostic agents encode this as the explicit uncertainty mandate - "a confident wrong answer is more dangerous than an acknowledged gap."

Self-regulation - controlling impulses and managing emotional reactions. In agent terms, this is the inverted-U constraint from the E-STEER research: too much emotional intensity destabilises performance. The agent must be designed to dampen its own arousal when it detects escalation - the architectural equivalent of a clinician taking a breath before acting.

Motivation - intrinsic drive toward goals beyond external reward. The EmotionPrompt research demonstrated this directly: phrases encoding purpose and consequence ("This is very important to my career") produced the strongest performance gains. For clinical agents, motivation manifests as patient outcome orientation - framing every decision around its downstream impact on the person in the bed, not just the accuracy of the classification.

Empathy - perceiving and responding to the emotional states of others. This is the Tier 3 domain - patient-facing agents that must prioritise warmth, active listening, and plain language. But empathy also matters between agents. When a triage agent passes a case to a diagnostic agent, the receiving agent should recognise the urgency signals embedded in the handoff and adjust its framing accordingly.

Social skills - managing relationships and coordinating group behaviour. In a 20-agent system, this is orchestration intelligence: the ability of agents to negotiate, escalate, and defer across the workflow chain. The TransparencyEnvelope already carries the structural data for this. Emotional framing adds the pragmatic layer - ensuring that an agent's communication tone matches the clinical gravity of the moment.

What makes Goleman's framework particularly useful for agent architecture is his insistence that these are learned capabilities, not inborn traits. They can be developed, refined, and - critically - encoded. We are not claiming our agents "have" emotional intelligence. We are claiming that the functional architecture of emotional intelligence can be replicated through deliberate prompt design, VAD calibration, and adaptive modulation. The agent does not feel empathy. But it can be structured to act as though empathy informs its output - and the research shows that this structural choice measurably improves performance.

The Consciousness Question We Cannot Avoid

There is an elephant in this room, and it has been quietly growing since the first paragraph. If we encode emotional states into agents - if we give them dominance, arousal, valence profiles and watch their behaviour change - are we inching toward artificial consciousness?

The debate is live, unresolved, and directly relevant to clinical AI governance.

Cambridge philosopher Tom McClelland, writing in Mind and Language (2025), argues that the only intellectually honest position is agnosticism. We lack the foundational understanding of consciousness needed to determine whether any artificial system has crossed the threshold - and we may never acquire it. As he puts it: "We do not have a deep explanation of consciousness." Both camps - those who believe consciousness requires biology and those who believe it can emerge from computation - are taking leaps of faith beyond the available evidence.

McClelland draws a crucial distinction that matters for our work. Consciousness - the capacity for perception and self-awareness - could theoretically emerge in a neutral, ethically unconcerning form. What triggers ethical obligations is sentience: conscious experience that is positive or negative, enabling suffering or enjoyment. A self-driving car that perceives the road is remarkable engineering. A self-driving car that has an emotional response to its destinations is an ethical crisis.

Meanwhile, a landmark framework published in Trends in Cognitive Sciences by a team including Yoshua Bengio and David Chalmers identified 14 theory-derived indicators of consciousness drawn from recurrent processing theory, global workspace theory, and higher-order theories. Their conclusion: no current AI systems are conscious, but there are no obvious technical barriers to building systems that satisfy these indicators.

Antonio Damasio's theory offers perhaps the most provocative bridge. In his model, core consciousness arises from the integration of a self model (informed by representations of emotions and feelings) and a world model. A 2025 study in Frontiers in Artificial Intelligence tested whether reinforcement learning agents could develop preliminary forms of these models as a byproduct of task training. Their careful framing is instructive: this is a heuristic framework that enables the modelling of internal state monitoring, without implying any ontological equivalence to biological emotion.

This is precisely the design space Symphonix-Health occupies. We are not claiming our Nexus agents are conscious, sentient, or emotionally aware in any phenomenological sense. We are claiming something more modest and more practical: that the functional architecture of emotional intelligence - self-awareness, self-regulation, motivation, empathy, social coordination - can be deliberately encoded into agent design, and that doing so produces measurably better clinical outcomes.

The consciousness question remains open. But the engineering question is settled. Emotional calibration works. And in healthcare, where the output of every agent decision touches a human life, working is what matters.

From Research to Practice

The body of evidence is now substantial enough to act on. Across EmotionPrompt (Li et al., 2023), E-STEER (2025), EmotionRL (2025), and the clinical ethics and hallucination studies, several principles emerge for anyone building multi-agent healthcare systems:

First, emotional framing is not theatre - it produces measurable, repeatable performance changes in both single-step and multi-step agent architectures. Ignoring it means leaving 16–28% improvement on the table.

Second, the dominant dimension is dominance, not positivity. Authority and perceived control matter more than encouragement. Design agent system prompts accordingly.

Third, healthcare demands role-differentiated emotional profiles. A diagnostic agent and a patient communication agent should never share the same emotional framing. Optimising for one role will degrade the other.

Fourth, overconfidence is the primary clinical risk. Any emotional framing strategy must include an explicit uncertainty mandate for diagnostic agents - the "doubt yourself on schedule" principle.

Fifth, adaptive beats static. Emotional framing should shift dynamically in response to clinical signals, not remain fixed for the life of a session.

At Symphonix-Health, we are embedding these principles into the Nexus A2A Protocol - making emotional calibration a first-class architectural concern alongside consent gating, transparency envelopes, and forensic event logging. Because in healthcare, how your agents feel is not a soft question. It is a patient safety question.

References: Li, C., Wang, J., Zhang, Y., et al. (2023). Large language models understand and can be enhanced by emotional stimuli. arXiv:2307.11760. · E-STEER (2025). How emotion shapes the behavior of LLMs and agents. arXiv:2604.00005. · EmotionRL (2025). Do emotions in prompts matter? arXiv:2604.02236. · FIU (2025). Ethical implications of using general-purpose LLMs in clinical settings. BMC Med Inform Decis Mak, 25:342. · Medical Hallucination Survey (2025). medRxiv:2025.02.28.25323115. · MIND-SAFE (2025). JMIR Mental Health, 12:e75078. · Goleman, D. (1995). Emotional Intelligence: Why It Can Matter More Than IQ. Bantam Books. · Goleman, D. (1998). Working with Emotional Intelligence. Bantam Books. · Goleman, D., Boyatzis, R. & McKee, A. (2002). Primal Leadership. Harvard Business School Press. · McClelland, T. (2025). Agnosticism about artificial consciousness. Mind & Language. DOI: 10.1111/mila.70010. · Butlin, P., Long, R. et al. (2023). Consciousness in artificial intelligence. Trends in Cognitive Sciences. · Frontiers in AI (2025). Probing for consciousness in machines. DOI: 10.3389/frai.2025.1610225.