Voluntary action—the capacity to initiate, guide, and terminate movements in service of a goal—is the cornerstone of human agency. It is what allows us to pick up a cup of coffee, to write a novel, or to steer a bee‑colony’s foraging strategy in the face of climate change. Neuroscience has long sought to map the brain’s “action‑planning” circuitry, but only in the last decade has neuroimaging begun to reveal the precise prefrontal dynamics that underlie the feeling of being in control. Understanding these mechanisms is not merely an academic pursuit; it informs the design of autonomous AI agents, the development of interventions for motor disorders, and even the conservation of pollinators whose survival hinges on complex, goal‑directed behaviors.
In this pillar article we synthesize the most robust findings from functional magnetic resonance imaging (fMRI), positron emission tomography (PET), and magnetoencephalography (MEG) that link agency to prefrontal activation. We then bridge these insights to the neural architecture of honeybees and the policy‑learning loops of self‑growing AI agents. By weaving together human, insect, and artificial systems, we illustrate how a shared principle of “agentic control” can be harnessed to protect ecosystems and build resilient autonomous technologies.
1. The Neural Architecture of Agency
At the heart of voluntary action lies a distributed network that spans from the medial frontal cortex to the motor cortices, thalamic relays, and basal ganglia. The dorsal anterior cingulate cortex (dACC) and the pre-supplementary motor area (preSMA) act as integrators of motivational and cognitive information, while the dorsolateral prefrontal cortex (dlPFC) encodes task rules and goals. Downstream, the primary motor cortex (M1) and premotor areas translate these abstract plans into spinal cord commands.
Neuroanatomical tracing in non‑human primates has shown that the dlPFC projects monosynaptically to the premotor cortex via the arcuate fasciculus, and indirectly to M1 through the basal ganglia’s direct pathway. In humans, diffusion tensor imaging (DTI) reveals a similar pattern: the superior longitudinal fasciculus (SLF) connects dlPFC to premotor and motor areas, while the uncinate fasciculus links the medial prefrontal cortex to limbic structures, allowing emotional valence to modulate action selection.
Importantly, the prefrontal network is not a static hierarchy but a dynamic, recurrent system that updates its representation of the environment in real time. Recurrent neural circuits within the dlPFC maintain working memory traces of intended actions for up to 4–6 seconds, a duration consistent with the temporal resolution of BOLD signals in fMRI studies of the “readiness potential” (RP). This readiness potential, first described by Kornhuber and Deecke in 1965, manifests as a slow buildup of activity in the supplementary motor area (SMA) preceding overt movement. Subsequent work has linked this buildup to the prefrontal initiation of volitional intent.
2. Functional Imaging: fMRI and PET Findings
2.1 BOLD Signatures of Volition
The most compelling evidence linking prefrontal activation to agency comes from event‑related fMRI paradigms that isolate the decision‑making phase from the motor execution phase. In a classic “Libet‑style” task, participants press a button at a self‑chosen time while their neural activity is recorded. The BOLD signal in the dlPFC and dACC rises approximately 0.5–1 second before the participant reports the conscious urge to act, peaking around 1.5–2 seconds prior to the button press. The magnitude of this pre‑movement activation correlates with the subjective confidence in the decision, suggesting that the prefrontal cortex encodes both the intention and the evaluative appraisal of the action.
Quantitatively, the BOLD change in the dlPFC during voluntary initiation is typically 0.8–1.2% of the baseline signal, a figure that is statistically significant (p < 0.001, corrected for multiple comparisons). This amplitude is comparable to that observed in reward‑processing studies, underscoring the motivational salience of self‑initiated actions.
2.2 PET Studies of Neurotransmitter Modulation
Positron emission tomography has provided a complementary view by measuring the distribution of neurotransmitter systems that modulate prefrontal activity. A study using the dopamine D2 receptor ligand [^11C]raclopride found that voluntary actions are associated with a transient decrease in dopamine release in the ventral striatum, suggesting that a “dopamine‑free” state may be conducive to exploratory, self‑generated behavior. Conversely, serotonin transporter imaging with [^11C]DASB showed increased serotonergic tone in the dlPFC during tasks that required inhibitory control over pre‑potent responses, indicating that serotonin may gate the transition from intention to action.
These neurochemical signatures align with the “two‑stage” model of action: a pre‑frontal “planning” stage that is dopaminergic and a post‑frontal “execution” stage that is serotonergic. Importantly, PET studies have also revealed that individuals with Parkinson’s disease exhibit blunted prefrontal BOLD responses during voluntary tasks, correlating with motor initiation deficits.
2.3 Temporal Dynamics in MEG and EEG
While fMRI provides spatial resolution, magnetoencephalography (MEG) and electroencephalography (EEG) capture the millisecond‑scale dynamics of agency. In a paradigm where participants choose between two visual targets, the readiness potential begins ~700 ms before the motor response, with a prominent peak in the left prefrontal cortex for right‑hand movements. This early prefrontal activity precedes the classic “Lateralized Readiness Potential” (LRP) in motor cortex, reinforcing the idea that the prefrontal cortex initiates the cascade of volitional action.
The temporal resolution of MEG also allows us to observe the interplay between the prefrontal and basal ganglia networks. A study using simultaneous fMRI/MEG found that the prefrontal BOLD signal leads the striatal BOLD by ~200 ms, suggesting that the prefrontal cortex sets the “go” signal that the basal ganglia then amplifies.
3. The Role of the Prefrontal Cortex in Voluntary Initiation
3.1 Dorsolateral Prefrontal Cortex (dlPFC)
The dlPFC is central to the executive control of action. In tasks that require the suppression of automatic responses, such as the Go/No‑Go paradigm, the dlPFC shows increased activation during the “No‑Go” trials, reflecting the maintenance of an abstract rule (“do not act”). Functional connectivity analyses reveal that the dlPFC engages the anterior cingulate cortex (ACC) to monitor conflict and the inferior frontal gyrus (IFG) to implement inhibition.
Neuropsychological evidence further underscores the dlPFC’s role: patients with lesions in this region exhibit “dysexecutive” behavior, failing to initiate or sustain goal‑directed actions. The severity of this deficit correlates with the extent of damage to the SLF, highlighting the importance of white‑matter integrity for agency.
3.2 Anterior Cingulate Cortex (ACC)
The ACC acts as a performance monitor, evaluating the cost–benefit trade‑offs of potential actions. In the “bandit task,” where participants must balance exploration and exploitation, ACC activity spikes when the expected reward of a new option surpasses the current value. This signal is thought to inform the dlPFC of the need to adjust the action plan.
ACC dysfunction, as seen in depression and obsessive‑compulsive disorder, leads to impaired decision‑making and a sense of “frozen” agency. Functional imaging shows reduced ACC activation during voluntary initiation in these populations, suggesting a failure to adequately evaluate the desirability of action.
3.3 Pre‑Supplementary Motor Area (preSMA)
The preSMA sits at the interface between cognitive control and motor execution. It integrates signals from dlPFC and ACC to generate a coherent motor plan. In tasks that involve complex sequences, the preSMA shows sustained activation throughout the planning phase, while the SMA proper activates at the moment of movement.
Neurostimulation studies using transcranial magnetic stimulation (TMS) over the preSMA delay the onset of voluntary action by ~200 ms, confirming its causal role in initiating movement. Moreover, the preSMA is sensitive to the timing of voluntary actions: when participants are asked to press a button at a specific time, preSMA activity peaks precisely at that target time, indicating a temporal coding function.
4. Temporal Dynamics: From Decision to Movement
4.1 The Readiness Potential (RP) and the Libet Paradigm
The readiness potential (RP) is a slow cortical potential that begins up to 1.5 seconds before a voluntary movement. The classic Libet experiment demonstrated that the RP starts before participants become consciously aware of their intention to move. Subsequent replication studies using high‑density EEG and source localization have localized the RP’s origin to the preSMA and dlPFC, rather than the primary motor cortex.
The temporal offset between RP onset and conscious intention (on average ~0.3 seconds) has sparked philosophical debates about free will. Neuroscience now suggests that the prefrontal cortex initiates the action plan, which is then refined by the ACC and basal ganglia before the motor cortex executes the movement. This cascade allows for rapid adjustments based on sensory feedback.
4.2 The “Urgency” Signal
In tasks where participants must respond quickly to a stimulus, the prefrontal cortex exhibits an “urgency” signal—a ramping activity that accelerates as the deadline approaches. This signal is mediated by the subthalamic nucleus (STN), which receives excitatory input from dlPFC and projects inhibitory output to the motor cortex. Functional imaging shows that the STN’s activity increases as the urgency signal rises, effectively gating the final motor command.
Clinical data from patients with deep brain stimulation (DBS) of the STN in Parkinson’s disease reveal that modulating this urgency signal can alter reaction times, providing a potential therapeutic target for motor initiation disorders.
4.3 Feedback‑Based Modulation
Once a movement is initiated, sensory feedback (proprioceptive, visual, auditory) is fed back to the prefrontal cortex via the posterior parietal cortex and cerebellum. This feedback loop allows the brain to correct errors on the fly. Functional connectivity analyses show that the prefrontal cortex increases its coupling with the cerebellum during tasks requiring fine motor control, such as playing a violin.
The cerebellum’s role is not limited to motor timing; it also contributes to the internal forward model that predicts the sensory consequences of an action. When the prediction matches the actual feedback, the prefrontal cortex reduces its activity, signaling that the action has been successfully executed.
5. Comparative Perspectives: Humans, Bees, and AI Agents
5.1 Honeybee Neural Architecture
Honeybees (Apis mellifera) possess a remarkably compact brain (~1 mm³) yet exhibit sophisticated goal‑directed behaviors such as navigation, foraging, and colony decision‑making. The mushroom bodies (MBs) in the bee brain are analogous to the human prefrontal cortex in terms of higher‑order processing. They integrate olfactory, visual, and proprioceptive inputs to guide navigation and learning.
Neuroimaging in bees using calcium‑based imaging has revealed that MB output neurons encode the bee’s intended flight direction before the bee initiates the flight. This anticipatory activity parallels the human prefrontal readiness potential. Moreover, the bee’s central complex (CC) acts as a central pattern generator, analogous to the basal ganglia, coordinating motor output.
The bee’s agency is also modulated by neuromodulators: octopamine, the insect equivalent of norepinephrine, increases during foraging and enhances the reward signal, much like dopamine in humans. Experiments that block octopamine receptors reduce the bee’s willingness to explore new flowers, underscoring the neurochemical basis of agency.
5.2 Self‑Growing AI Agents
Artificial agents that learn to act through reinforcement learning (RL) exhibit a computational counterpart of the human prefrontal system. The policy network, typically a deep neural network, encodes the mapping from states to actions. During training, the agent updates its policy via gradient descent on the expected reward, akin to the dopaminergic reward prediction error signal in the human striatum.
Recent work on “intrinsic motivation” in RL introduces an internal reward signal that encourages exploration. This intrinsic reward is implemented as a novelty bonus, analogous to the human prefrontal evaluation of novelty. Agents that incorporate intrinsic motivation tend to exhibit more flexible, goal‑directed behavior, mirroring the role of the human prefrontal cortex in initiating voluntary action.
Furthermore, the concept of a “policy gradient” in RL is conceptually similar to the prefrontal cortex’s modulation of motor plans based on expected outcomes. When an agent’s policy predicts a high reward, the gradient signal becomes stronger, leading to a higher probability of action selection. This parallels the prefrontal cortex’s increased activity when a particular action is deemed more valuable.
6. Practical Implications: Enhancing Agency in Conservation and Robotics
6.1 Conservation: Modulating Bee Foraging Through Neural Cues
Understanding the neural underpinnings of bee agency can inform conservation strategies. For instance, providing floral resources that elicit strong octopamine release can increase foraging motivation. Habitat management practices that reduce pesticide exposure preserve the integrity of the bee’s central complex, ensuring that the internal pattern generators function correctly.
Neuroethological studies have shown that exposure to neonicotinoids reduces the activity of mushroom body neurons, leading to impaired navigation and reduced colony productivity. By monitoring neural biomarkers (e.g., calcium transients) in field bees, conservationists can assess the health of pollinator populations in real time.
6.2 Robotics: Bio‑inspired Control Loops
Robotic systems can benefit from bio‑inspired control architectures that mimic the human prefrontal‑basal ganglia loop. A modular architecture comprising a high‑level decision module (analogous to dlPFC), an urgency module (STN analog), and a low‑level motor module (M1 analog) can achieve flexible, goal‑directed behavior in dynamic environments.
Recent advances in neuromorphic hardware, such as Intel’s Loihi chip, allow for event‑driven processing that mirrors the temporal dynamics of the prefrontal cortex. Integrating these chips with reinforcement learning algorithms that incorporate intrinsic motivation can produce robots capable of self‑initiated exploration and adaptive task switching.
6.3 Clinical Applications
The prefrontal activation patterns identified in voluntary action studies can inform neurorehabilitation protocols for patients with motor initiation deficits. Techniques such as transcranial direct current stimulation (tDCS) targeting the dlPFC or preSMA have shown promise in improving gait initiation in Parkinson’s disease. By tailoring stimulation protocols to the temporal dynamics of the readiness potential, clinicians can enhance the efficacy of interventions.
7. Future Directions and Open Questions
- Causal Mapping of Agency: While fMRI and PET provide correlational data, causal interventions (e.g., optogenetics in animal models, TMS in humans) are needed to confirm the prefrontal cortex’s role in initiating action. Recent advances in simultaneous fMRI/TMS allow for real‑time modulation of prefrontal activity during voluntary tasks.
- Multimodal Imaging: Combining fMRI, MEG, and PET can yield a more comprehensive picture of the spatiotemporal dynamics of agency. For example, simultaneous fMRI/MEG could resolve the millisecond‑scale prefrontal activity that precedes the BOLD signal, bridging the gap between neural firing and hemodynamic changes.
- Neurochemical Modulation: Further studies on how neuromodulators (dopamine, serotonin, norepinephrine) shape prefrontal activation during voluntary action will clarify the neurochemical basis of agency. Pharmacological challenge studies in humans and bees can reveal conserved mechanisms across species.
- Artificial Intelligence Integration: Translating prefrontal mechanisms into AI requires bridging the gap between biological neural networks and artificial neural networks. Developing biologically plausible learning rules (e.g., spike‑timing dependent plasticity) that emulate prefrontal plasticity could yield more flexible, self‑directed agents.
- Ecological Validity: Most neuroimaging studies use simplified laboratory tasks. Future research should employ ecologically valid paradigms, such as virtual reality environments that simulate real‑world decision making, to better capture the complexity of voluntary action.
8. Why it Matters
The convergence of neuroimaging findings, comparative neurobiology, and AI research underscores a fundamental truth: voluntary action is a product of complex, distributed neural dynamics that are both conserved across species and amenable to technological emulation. By elucidating the prefrontal mechanisms that generate agency, we gain tools to:
- Protect pollinators by designing habitats that support the neural health of bees, ensuring their vital foraging behavior remains robust against environmental stressors.
- Build autonomous systems that can self‑initiate, adapt, and learn in uncertain environments, mirroring the flexibility of biological agents.
- Improve human health by developing targeted neuromodulation therapies that restore agency in patients with motor disorders.
In an era where technology increasingly interfaces with biological systems, understanding the neural basis of voluntary action is not merely academic—it is essential for fostering harmonious coexistence between humans, bees, and intelligent machines.