We studied two queenright experimental honey bee colonies of mixed European subspecies (Apis mellifera subsp.), each containing approximately 3,000 workers plus brood. These bees were marked (see below) as part of another project (McHenry et al., 2025a) but analyzed here for our investigation. Briefly, it was important that bees be individually identifiable from the onset of the experiment since we were interested in tracking the flow of information from bee to bee via waggle dance communications. To that end, we marked each cold-anesthetized worker with a unique identifier (UID). Each UID consisted of a combination of a colored, numbered disc tag (BetterBee, Greenwich, NY) that we glued to the thorax and a colored stripe from a paint pen (POSCA, USA) applied dorsally to the abdomen. We used sets of disc tags numbered 1–100 in five colors and paint pens in six colors to yield 3,000 distinct UID values. We placed each marked bee in a recovery chamber with the queen and access to sucrose solution until we were done with marking. In all, we marked approximately 3,000 workers from each colony. We installed workers in an observation hive with the queen, usually by mid-afternoon on the second day. We created and worked sequentially with two fully-marked colonies.

Each observation hive consisted of three American Standard Deep frames arranged vertically in a wooden scaffold such that the comb surfaces aligned in a single plane. This setup was enclosed with transparent Lexan, allowing unobstructed visual access to activity on the comb surface, including waggle dances. To encourage waggle dancing, we selected frames with some empty cells available for nectar storage, as previous work has shown that shortage of storage space leads to increased forager wait time for nectar unloading, which then reduces the likelihood of dances Seeley, 1989; Seeley and Tovey, 1994; Seeley, 1995). We housed the observation hives indoors at the Prices Fork Research Center in Blacksburg, Virginia, with hives connected to the outdoors via a plastic tube measuring approximately 5 cm x 30 cm.

We collected field data from July 11, 2022, to July 23, 2022, on warm sunny days with favorable foraging conditions. High summer tends to be a preferable time to train bees to feeders in our study area, since there is a relative scarcity of floral resources in the landscape, making our training sugar solution more attractive (Seeley et al., 1991; Ohlinger et al., 2022). Colonies received supplemental feeding during periods of poor weather, but this supplemental food was removed prior to data collection.

For each experimental colony, we trained bees to forage at an artificial feeder station using previously established stepwise training protocol (Schürch et al., 2013; Couvillon et al., 2015; Schürch et al., 2016, 2019; Ohlinger et al., 2022; Couvillon et al., 2023; McHenry et al., 2025a; McHenry et al., 2025b). Briefly, the training occurred in two phases called the pre-training phase and experimental phase. In the pre-training phase, we trained an initial cohort of foragers to collect 2 M sucrose solution scented with lavender (5 μL/L) from our artificial feeder station. This initial pre-training cohort would become our “origin” bees, so named because they would go on “originate” their own waggle dance communication networks during the experimental phase.

We initiated pre-training by loading a Petri dish with a few drops of the sucrose solution and placing it inside a small box, which we then secured over the exterior hive entrance using a tripod and duct tape. This setup ensured that foragers exiting the colony would explore the box, and that some would find the food source. A transparent window in the top of the box allowed us to monitor the Petri dish for drinking foragers. When at least three individual bees were drinking from the Petri dish, we carefully dislodged the box and gently transferred the dish directly onto the tripod platform, still situated just outside the hive entrance, taking great care not to disturb drinking bees. When 3–5 individual bees had each completed at least three foraging trips to and from the Petri dish, we swapped in the full feeder setup: a 4 oz glass jar (approximately 29.6 ml; Ball, Westminster, CO, USA) mounted on a plastic base with collection wells situated around the perimeter of the jar’s mouth. The tripod base’s platform was covered with a blue plastic tablecloth and its height set at 1 meter above the ground.

We then gradually relocated the feeder farther away from the colony in a stepwise fashion, moving across an open field in increments of 10–15 meters at a time. At each step, we waited until 5–15 bees had made at least three return trips, confirming that they had learned and remained committed to the current feeder location, before advancing to the next step. This stepwise procedure continued for 10–12 steps, until we reached the initial experimental feeder distance – 215 meters in Trial 1 and 315 in Trial 2 (see below for feeder distance rationale). This process typically lasted through the late afternoon or early evening. Allowing bees to accumulate multiple rewarding visits at the final feeder location increased the likelihood that they would return the following day as origin bees and perform waggle dances to recruit nestmates (Seeley, 1995).

To begin the experimental phase, we re-established the feeder at the final location reached during pre-training. The experiment began when the first bee from the pre-training cohort returned to the feeder, which was freshly replenished with unscented 2 M sucrose solution. Such a concentrated solution provided a highly rewarding foraging experience for returning bees, which encouraged them to advertise the feeder and recruit a new wave of bees, which were our experimental phase recruits. These newly-recruited bees were of course already marked, but naïve to the location of our feeder. We used unscented sucrose solution to ensure that they would need to rely as much as possible on waggle dance information rather than on olfactory cues (Tautz and Sandeman, 2003; Tautz, 2022; Tautz, 2023).

Since one of our interests was to examine the role of individual waggle dance calibrations (Schürch et al., 2013; McHenry et al., 2025a) in the structure and propagation of the recruitment network, it was important to capture waggle dance data for as many experimental phase bees as possible. This is because, if a bee did not dance, we could not obtain a calibration value for her, and she would be ineligible for this portion of the analysis. To that end, we allowed foraging and recruitment to continue until our origin bees and 10–20 experimental phase recruits had each completed at least 2–3 unique waggle dance bouts advertising the feeder location. Four researchers participated in experimental phase data collection: two monitored and recorded foraging visits at the feeder, while two others observed and filmed waggle dances at the observation hive, one on either side of the colony so as to monitor both sides of the dance floor at once. We logged all feeder visits and waggle dances using ODK Collect on Samsung tablets, recording each bee’s UID and the event’s timestamp.

We filmed the bottom frame, or “dance floor,” of each observation hive continuously during experimental phase data collection, using two Canon Vixia G50 camcorders mounted on tripods. We placed one on either side of the observation hive, positioned approximately 1 meter from the comb surface to reduce parallax. Video data were recorded at 30 frames per second using the 3840x2160 resolution setting, and stored on SanDisk Extreme 256 GB SD cards. We labeled and stored SD cards as original video data, but also backed up all video data to a cloud-based repository and an external storage device (Seagate One Touch with Hub, 8 TB).

We had two main goals for our video data analysis: (1) to match each experimental phase recruit with the dancer who recruited her to map out the recruitment network, and, whenever possible, to (2) obtain dancers’ calibrations to study their effect on the network as a node attribute. Detailed description of methods used for video data extraction may be found in the Supplementary Online files for McHenry et al. (2025a). Briefly: to achieve this first goal, we matched each experimental phase recruit with the dancer from whom she learned of the feeder’s location (recall that this dancer could have been an origin bee or a fellow experimental phase recruit). To do this, we searched video data in the minutes preceding each bee’s first recorded feeder visit in DaVinci Resolve 18 (Version 18.0.4, Build 5) to find the last waggle dance where she followed at least one waggle run before arriving at the feeder (Seeley et al., 2000; Couvillon et al., 2012). Our criterion for dance following was being within antennal contact distance with the dancer during a waggle run (Rohrseitz and Tautz, 1999; Tanner and Visscher, 2008; Grüter and Ratnieks, 2011; Kietzman and Visscher, 2019).

To achieve the second goal of extracting waggle dance calibration values, we followed the dance decoding methods of Couvillon et al. (2012) in DaVinci Resolve, decoding four non-first, non-last waggle runs with two left-lead-ins and two right-lead-ins for each bee. The arithmetic mean of these four waggle runs gives a representative estimation of that bee’s waggle dance distance calibration (Couvillon et al., 2012; Schürch et al., 2013, 2019). Whenever possible, for each dancer-follower dyad, we decoded dancer waggle runs from the same dance bout that the successful recruit herself followed prior to arriving at the feeder, to represent as closely as possible the information that the recruit had received. For the successful recruits, we selected waggle runs from the first dance where our waggle dance decoding criteria could be met, and which occurred after the recruit had made at least three consecutive feeder visits (Chatterjee et al., 2019). This yielded waggle run duration values, measured in seconds, for each member of a successful dancer-follower dyad.

Here, we had to contend with a tricky feature of this dataset: for the purposes of the original study (McHenry et al., 2025a), waggle dance data were collected at not one, but two feeder distances from the colony as mentioned above (see that study’s Supplementary Online for rationale and details). This posed a challenge for our analysis, since calculating differences between calibrations in dancer-follower dyads would not be appropriate if waggle dances were advertising different distances. To account for this, we scaled all bees’ waggle run durations according the universal waggle dance calibration data object (Schürch et al., 2019) for the appropriate experimental distance (either 215 m, or 315 m). We first separated our observed waggle run duration data by distance category, and then joined them to a data frame containing a set of 1000 randomly selected, simulated waggle run duration values from the universal waggle dance calibration data object (Schürch et al., 2019) for the appropriate experimental distance. We then conducted scaling on each distance category’s full dataset including simulated and observed values before re-extracting our now-scaled observed values. For our analysis, we used these scaled duration values as calibration values.

The field of network analysis has its own vocabulary that is not commonplace even in closely related fields. Here ( Table 1 ) we briefly define some key terms that we use to describe our networks and analyses.

To construct the waggle dance communication network, we used the dancer-follower relationship datasets obtained from video analysis (Goal 1, see above) to create a network edge list using the igraph R package (Csárdi and Nepusz, 2006; Csárdi et al., 2024). We represented recruitment relationships as directed edges, with the edge pointing from dancer to follower along with the flow of information. Each node representing an experimental phase recruit thus has one inbound tie representing the receipt of feeder location information from the dancer that recruited her, but can have any number of outbound ties depending on the number of nestmates she herself went on to recruit. We also added isolate nodes to the network, which we considered to be origin bees that continued visiting the feeder during the experimental phase but did not recruit any nestmates to it, whether from lack of dancing or lack of dancing success. We excluded from the network any bee that participated in the pre-training phase but did not continue visiting the feeder after the start of the experimental phase.

Network density is defined as the proportion of observed ties relative to the maximum number of ties possible for that network, which is also referred to as a complete graph (i.e. if every node were connected to every other node) (Bang-Jensen and Gutin, 2018). Density is a common network descriptive statistic, giving an overall measure of the level of connectivity or sparseness of a network. However, we considered that the case of a complete graph as a baseline comparison to our observed number of ties would be biologically meaningless, as there is no real-world scenario where every bee recruits every other bee to the same food source. Rather, a biologically-informed complete graph (hereafter referred to as BICG) of honey bee foraging would be one where every dancing bee had a perfect record of successfully recruiting every nestmate for whom she performed a waggle dance.

We therefore set out to simulate a BICG as an empirical estimate of an upper bound on network edges for a honey bee recruitment network using a deterministic compartmental model (DCM) of the susceptible-infected (SI) model family (Anderson and May, 1991; Gernat et al., 2018). Traditionally formulated deterministic SI models feature individuals that are in one of two states: either “susceptible” or “infected,” and an infected individual may infect susceptible individuals when they come in contact based on a given probability value that is specified in the model. We reimagined an SI model in the context of honey bee waggle dance communication, wherein “infected” bees are those with knowledge of a food resource, “susceptible” bees are nestmates that follow dance communications, in part because they are currently unemployed foraging elsewhere, and an “infection” would represent a successful recruitment to the resource of a susceptible bee by an infected bee. To simulate the BICGs of honey bee foraging, we set the “infection” probability to 1. The resulting network simulation therefore represents what the recruitment network might look like if bees were 100% effective at recruiting every nestmate they contacted while dancing – a more biologically relevant and informative upper bound that we then used in place of a complete graph in calculating network density.

Since density is calculated with respect to the complete graph for the set of nodes and edges belonging to the observed network, we parameterized and simulated a BICG for both trials. To do this, we used the dcm() function from the R package EpiModel (Jenness et al., 2018). We specified the number of initially-infected individuals for each trial as the number of origin bees that made at least one feeder visit during the experimental phase, since these were bees that had knowledge of the feeder location and thus had the opportunity to advertise it. We set the total population to 1,000 individuals, since colonies were created with 3,000 marked bees and roughly one-third of the workers within a colony may be involved in foraging activities at any given time (Seeley, 1995, Seeley, 2019).

Here we calculated classical density for each trial using both conventional complete graphs and corrected density using our BICG as upper bounds. For the sake of comparability between classical and BICG (corrected) density values, we first added dummy isolate nodes to our observed networks to get the node total to 1,000. We then supplied that graph object to the edge_density() function from the igraph package to obtain classical density for each trial. Then we calculated a corrected density for each trial by dividing the number of edges in the observed graph by the number of edges in the BICG.

Burstiness is a concept used to describe the temporal clustering of events in a time series. In the context of network analysis, it refers to the phenomenon where nodes exhibit alternating periods of high and low activity (Barabási, 2005). This pattern indicates that events or interactions occur in clusters, and it deviates from a uniform or random distribution of actions over time.

To explore burstiness in waggle dance behavior, we first created a timeline of waggle dance bouts for each bee. To do this, we used the waggle dance data collected at the observation colony during the experimental phase to calculate the time interval between each successive dance for each bee, excluding overnight periods when data collection paused. A bee was eligible for burstiness analysis if she performed at least two waggle dances during the experimental phase, since the burstiness coefficient is calculated based on inter-event times. Therefore, we extracted the timelines of only bees who met these criteria, and we calculated the coefficient for each and then plotted the distribution of burstiness values.

We set out to test whether individuality in waggle dance calibrations influenced the structure of waggle dance communication networks. Previously, we showed that individual calibrations (Schürch et al., 2016), rather than causing miscommunication, supported recruitment success depending on the relationship between the dancer’s and follower’s calibrations. Specifically, recruitment is more likely to succeed when a follower bee’s communicated waggle run duration for a given distance, or her calibration, is shorter than that of the dancer that recruited her (McHenry et al., 2025a). We call this type of calibration relationship the positive mismatch condition (dancer calibration – follower calibration > 0) and would likely result in the follower bee overshooting the target in the landscape, which might then double her resource discovery opportunities as she flies back over it on the way back to the hive (McHenry et al., 2025a). Now we wanted to know whether positive mismatch in calibrations would also affect the structure of the larger recruitment networks emerging from waggle dance interactions. Here, again, we needed to customize the analysis to reflect the biology of recruitment: since each bee could only be recruited to the feeder once, each bee could therefore have only one inbound tie, representing the successful receipt of feeder location information. To test the effect of calibration differences on network structure while respecting this one-inbound-tie network constraint, we implemented a custom node-label permutation test (Whitehead, 2008; Croft et al., 2009; Croft et al., 2011; Holt and Sullivan, 2023). In general, node-label permutation tests for significance by repeatedly shuffling the specified node attribute and calculating the outcome of interest over many iterations to generate a null distribution against which observed values are then compared. Here, we reshuffled scaled calibration values (described above) among nodes in our observed network and then calculated the number of ties in each permuted network that matched the longer-to-shorter calibration relationship criterion between dancers and followers. We then compared the observed number of ties matching this criterion against the permuted null distribution to obtain a p-value.

We carried out all data cleaning, variable calculation, analysis, and visualization in R version 4.4.2 (R Core Team, 2024) and the following R packages: igraph v. 2.0.3 (Csárdi and Nepusz, 2006; Csárdi et al., 2024), knitr v. 1.46 (Xie, 2014; Xie, 2015; Xie, 2024), rmarkdown v. 2.26 (Xie et al., 2018; Allaire et al., 2024), tools v. 4.4.1 (R Core Team, 2024). We used the package grateful v. 0.2.4 (Rodríguez-Sánchez and Hutchins, 2020) to facilitate proper citation of all R packages used.

Data and code underlying this manuscript are made accessible through the Virginia Tech Data Repository at DOI: 10.7294/29395904.

We calculated network density values for both trials using both a classically defined complete graph and our biologically-informed complete graph (BICG; see above). For Trial 1, the classical network density for the observed network was 3.60 x 10^-5 ( Figure 1 , Trial 1a), and the corrected network density calculated with our simulated BICG was 3.68 x 10^-2 ( Figure 1 , Trial 1b). For Trial 2, classical network density for the observed network was 1.50 x 10^-5 ( Figure 1 , Trial 2a), and corrected density using the simulated complete graph was 1.55 x 10^-2 ( Figure 1 , Trial 2b). In this way, both trials reveal that honey bee waggle dance recruitment networks are sparsely connected, and that the biologically-informed density values differed from the classical density values by a factor of 1000.

Of the 63 bees that performed waggle dances during the experimental phase to advertise our feeders, 53 bees performed more than one dance and were thus eligible for burstiness analysis.

Overall, bees danced steadily, not burstily. We calculated a mean burstiness coefficient (95% CI) of -0.05 (-0.12 – 0.015). We also compared our observed distribution of burstiness coefficients against a normal distribution by quantifying skewness and kurtosis. Our burstiness coefficient distribution’s skewness was -0.68, indicating a left-skew, and its kurtosis was 3.26.

The flow of information in the waggle dance recruitment network was facilitated by calibration differences. In our bees’ communication networks, 76% of recruitment events occurred between dyads where the dancer’s calibration was longer than the follower’s, with 23 ties out of 30 matching this criterion ( Figure 2 ). This frequency of longer-to-shorter ties was higher than would be predicted by chance based on the permuted null distribution (p = 0.008). As a result of this phenomenon, chains of recruitment generally showed a trend of successively lower calibration values.

We found that, while both the classical density and biologically-informed density of observed recruitment networks were sparse, biologically-informed density calculated using our BICG yielded network density values that were higher than classical density by three orders of magnitude ( Figure 1 ). Even when using our biologically-informed, simulated upper bound for waggle dance recruitment, the resulting network density values indicate an extremely sparse network structure, containing only up to 3.5% of the edges predicted by our BICGs ( Figure 1 ). In other words, our model estimates that each trial’s pool of origin bees recruited only up to 3.5% of the nestmates that they might have, if they had a perfect recruitment record.

The sparseness of the recruitment network may be related to the fact that transmission of information and successful recruitment relies upon many signal receiver characteristics, choices, and what we might call luck. It is a leaky pipeline. For successful recruitment to occur, a potential receiver bee must first be within perceptive range of a waggle dance (Drosopoulos and Claridge, 2005), usually 3–6 cell diameters or 18–27 mm from the dancer (Tautz and Rohrseitz, 1998; Kietzman and Visscher, 2019). Even if she is within perceptive range, she may not choose to follow the dance (Tautz and Rohrseitz, 1998; Okada et al., 2008; Grüter and Farina, 2009). If she does, she must closely observe multiple waggle runs (Judd, 1994, mean 8 runs; Grüter et al., 2008, 17; Menzel et al., 2011, 20-23; Wray et al., 2012, 15.5), without being bumped off-course by nestmates. Having done this, she must then weigh this social information against any private information that she may possess (Grüter et al., 2008; Grüter and Ratnieks, 2011; Grüter et al., 2013), to decide whether act on the public dance information received, or do something else. Approximately 75 – 88% of dance-following interactions result in reactivation to a known foraging site on the part of experienced foragers (Biesmeijer and Seeley, 2005) rather than recruitment to the advertised food source (i.e., their dance-following does prompt a change in volitional state, but towards the use of private information rather than public (Grüter et al., 2008; Grüter and Farina, 2009; Grüter and Ratnieks, 2011; Hasenjager et al., 2020)). Finally, the last step in this leaky pipeline: 60-80% of bees who do set out to find the advertised resource will fail to do so (Esch and Bastian, 1970; Gould et al., 1970; Seeley, 1983; Seeley and Visscher, 1988; Seeley and Towne, 1992; Riley et al., 2005; Seeley, 1995). Considering all this together, perhaps such sparse network densities are unsurprising.

This leaky pipeline of recruitment, and the sparseness of recruitment networks that we found, are both consistent with the decentralized model of division of labor, where behavior is coordinated locally by the interaction of an individual’s morphology, genotype, age, and past experience with that individual’s particular social environment (Hölldobler and Wilson, 1990; Fewell and Page, 1993; Calderone, 1998; Beshers and Fewell, 2001; Ravary et al., 2007; Robinson, 2009). Most of these “leaks” in the recruitment pipeline may then be attributed to variation between bees’ predispositions and experience that leads only a subset of bees all the way to successful recruitment given the particular social stimuli on the dance floor. The sparseness of recruitment networks that we found is consistent with this paradigm. Additionally, however, sparse recruitment networks are consistent with what is known, if not fully understood, about honey bee foraging and recruitment: that the waggle dance is a famously inefficient system. The last leak in this pipeline, where 60 – 80% of bees who try to locate an advertised resource fail to do so (Esch and Bastian, 1970; Gould et al., 1970; Seeley, 1983; Seeley and Visscher, 1988; Seeley and Towne, 1992; Riley et al., 2005; Seeley, 1995), we consider not to be an example of local control of division of labor, but an inefficiency in the waggle dance as a semiotic system. In other words, this scenario is not a failure of the social environment to produce a task switch in a given individual, but rather a task switch initiated and failed. Furthermore, this failure incurs the cost of the energy needed to complete the unrewarded flight. As the sparseness of recruitment networks that we found is consistent with decentralized control of division of labor as well as inefficiency recruitment, it would now be interesting to examine the relative contributions of these two principles to the form and function of recruitment networks.

Our results also highlight the importance of calculating and interpreting network metrics in the context of the behavior and biology of the organism under study (Makagon et al., 2012; Alves et al., 2023). The default baseline for calculating network density, a complete graph where every node is connected to every other node (Bang-Jensen and Gutin, 2018), could not occur in the context of honey bee behavior. A forager bee, once committed to a food source, could not then be recruited to it again by every other forager bee. The large difference we found between the default method versus our biologically-informed approach emphasizes that such system-specific customization of network analyses can help glean more relevant insights into biological systems.

In this study, the adaptation of the deterministic SI model to simulate maximally-efficient recruitment was an important piece of such system-specific customization of analysis. These kinds of models have been used previously to study the social network of trophallaxis in honey bees (Gernat et al., 2018; Sun et al., 2020, with the same dataset) modeling the spread of infection through the sharing of crop content. As we did, those authors also used a deterministic SI model to simulate a graph representing the upper-bound of infection rates within a trophallaxis network, parameterized with a likelihood of infection set to 1. Here, we further customized the parameterization of the model according to the biology of waggle dance communication and recruitment: we (1) specified directed edges, since dance information flows from dancer to follower, and (2) supplied a total number of individuals (1,000) based on what is known about the proportions of workers that may be engaged in foraging at any given time (approx. 1/3) (Seeley, 1995; Seeley, 2019), with our total observation colony size being 3,000. An important limitation, however, is that we only studied the recruitment network for one food source at a time (our feeder). In reality, the dance floor of a freely foraging honey bee colony may host a medley of waggle dance signals advertising a variety of resources (Seeley, 1995; Couvillon et al., 2014; Garbuzov et al., 2015), and one third of the workers are not simply waiting around to be recruited, as they are in our model. It would now be interesting to customize such a model to reflect more specifically the proportion of the workforce that is on the dance floor, otherwise unoccupied, and in such a volitional state as to amenable to being recruited. To our knowledge, this work is the first to apply deterministic SI models to the study of waggle dance recruitment networks, but we suggest that there are further opportunities to extend their usefulness in studying the flow of information in bee societies.

We found that honey bees tend to waggle dance at regular intervals. Burstiness values close to zero, such as we observed, indicate high inter-event homogeneity. In other words, once bees begin to dance, they keep up a regular rhythm of dance bouts. Importantly, however, the functional role of burstiness (or lack thereof, as we found) on the spread of information through the honey bee foraging workforce is as yet unknown. Burstiness as a feature of node behavior in networks can either increase or decrease information flow, depending on the context and constraints of the network in question (Karsai et al., 2011; Lambiotte et al., 2013; Rocha and Blondel, 2013; Bakhouya and Gaber, 2015; Evans et al., 2020; Unicomb et al., 2021; Doi et al., 2023). For instance, while burstiness is associated with slowed spread in some human communication networks (Karsai et al., 2011), burstiness in honey bee trophallaxis networks is associated with a speedier spread of a simulated pathogenic infection (Gernat et al., 2018). However, what role burstiness might play in the regulation of collective foraging and recruitment networks is still unknown (although see Campos et al., 2016).

It is important to note that, unlike in a real-world foraging landscape, the feeder provides a single, ad libitum source of higher quality food, which generates tight commitment to a feeder (Couvillon et al., 2015; Schürch et al., 2016, Schürch et al., 2019; Ohlinger et al., 2022; Couvillon et al., 2023). It is not surprising, then, that a feeder training scenario such as ours, with highly regular foraging visits, would also have highly regular dance bouts. During a natural foraging trip, however, bees must visit, probe, and collect nectar from many flowers (Ribbands, 1949; Goulson, 1999), that may be dispersed patchily (Beekman and Ratnieks, 2000; Goulson, 2000; Goulson, 2003; Dornhaus and Chittka, 2004; Dornhaus et al., 2006). The time required to fill a crop may therefore vary widely on successive foraging trips, which would then translate to a more irregular rhythm of waggle dance bouts. This is in stark contrast to the flow of a feeder training experiment, where bees can get into a tight rhythm, coming back and forth from the feeder every 3–4 minutes (Ohlinger et al., 2022). It is therefore possible that the low burstiness we found might be overstated compared to a natural foraging scenario. Even so, we suggest that future studies of burstiness in social behaviors may provide a new avenue for understanding the inner workings of the honey bee superorganism.

We found that individuality in waggle dance calibrations is positively associated with the propagation of the waggle dance recruitment network ( Figure 2 ). This finding is consistent with and builds upon previous work that tested the role of individuality on recruitment success (McHenry et al., 2025a). In the waggle dance, each forager translates the real-world distance she has flown into a waggle run duration, and she does this according to her own (Schürch et al., 2016), linear (Schürch et al., 2019) distance-to-duration calibration curve. As a result, different bees advertising the same location may produce different waggle run durations, and these individual calibrations can therefore lead to calibration mismatch between dancers and followers. If the dancer’s calibration < follower’s calibration, the follower would interpret a closer distance than intended by the dancer (negative mismatch condition) and would be likely to undershoot the resource. On the other hand, if the dancer’s calibration > follower’s calibration, the follower would interpret a farther distance than intended by the dancer (positive mismatch condition) and would be likely to overshoot the resource. The magnitude of this mismatch will also vary across dancer-follower dyads. In our previous study (McHenry et al., 2025a), we were surprised to find that low mismatch magnitude did not support recruitment success, but the positive mismatch condition did. At the dyad level, then, it is the relationship between calibrations in given dancer-follower pairs that matters, rather than any individual attribute of one of the communication partners, that impacts the likelihood of communication success. Consistent with this dyad-level effect, and in keeping with our predictions, we found that the propagation of the recruitment network was likewise supported by positive mismatch between dancers’ and followers’ calibrations (p = 0.008). In other words, the flow of information from bee to bee and the formation of the network was facilitated, not hindered, by individuality.

It would now be interesting to determine how variety in individual calibrations interacts with a honey bee colony’s collective foraging strategy. It is known that forager bees generally fall into one of two main categories related to foraging strategy and information use: scouts, who tend to rely on private information and search independently for floral resources on the landscape, dancing when they have found something profitable, and non-scouts (Liang et al., 2012), who tend rely on public information and wait for waggle dance information to direct them to a profitable resource (Seeley, 1983; Beekman et al., 2007; Liang et al., 2012; Grüter et al., 2013). A colony wishing to maximize their exploitation of the landscape by dispatching foragers efficiently should therefore have scouts who are effective recruiters, and non-scouts who are easy to recruit. As positive mismatch in calibrations is associated with recruitment success, we therefore wonder whether scout bees might have, on the whole, higher calibrations compared to non-scouts. In such a scenario, scout bees with higher calibrations would originate longer chains of recruitment ( Figure 2 ), where waggle dance information flows more readily down what we might call the calibration gradient.

Furthermore, there also likely exists a “sweet spot” in waggle dance calibration, where for a given resource distance, there is a duration that is just long enough to gain the benefit of positive mismatch, but not long enough to send recruits too far afield. A dancer whose calibration is too high compared to those of her nestmates would send would-be recruits too far, incurring greater flight costs and a wider search radius needed to locate the resource intended by the dancer (Tautz, 2022, 2023). We suggest that those bees who contribute the most to downstream propagation of waggle dance information would have calibration values somewhere in that sweet spot: high enough that a large proportion of followers would fit the positive mismatch condition, but not so high that mismatch magnitudes would be too large, sending most followers too far to find the resource. At the same time, the benefit of positive calibration mismatch between dancers and followers implies that there may be a limit on the length of recruitment chains. Information should flow more readily down the calibration gradient at first, from bees with higher calibrations to bees with lower calibrations, but only up to a point: eventually, the most recent recruits’ calibrations would be relatively low, and most of their followers would have the negative mismatch condition, which does not support recruitment. It would now be interesting to test whether bees with relatively high calibration values originate longer chains of recruitment, which ultimately peter out when bees at the end of the chain have sufficiently low calibrations. Overall, our findings emphasize that there is still more to learn about the role of individuality in the collective regulation of the honey bee colony superorganism (Hölldobler and Wilson, 2009; Borofsky et al., 2020).

Managing complexity is one of the greatest problems facing modern-day human organization (Beer, 1981; Espejo and Reyes, 2011). The honey bee colony has already inspired a wide variety of models and algorithms seeking to refine strategies to accomplish this goal in fields ranging from computing networks (Wedde et al., 2004, 2005; Chamoli et al., 2015; Sun et al., 2020; Doi et al., 2023), informatics (Beynon-Davies, 2010), and the management of big data with artificial intelligence (Foss and Espejo, 2018) to sustainable civil engineering (Sahai et al., 2025). However, lack of empirical knowledge about honey bee network structures and their real-world parameters has mostly restricted the bio-inspiration available to these algorithms and models to general roles associated with division of labor in a colony (Farooq, 2009). A critical concern in stream-based computing is achieving low latency, or temporal lag in data transmission, coupled with high throughput, or volume of data transmission (Sun et al., 2020). Recent attention has focused on the testing of parameterization conditions of bee-inspired algorithms and models, including various network density conditions, in order to achieve these goals (Bakhouya and Gaber, 2015; Verma et al., 2025).

As far as we are aware, our data are the first to offer density and burstiness parameterization guidance directly from in vivo bee networks. We therefore hope and anticipate that these results may be of interest to researchers wishing to test bio-inspired algorithm parameterizations. In particular, it would be interesting to test the behavior of stream-based computing models when parameterized according to the density and burstiness properties observed in real bee networks. Would bee-inspired parameterizations offer good latency and throughput, or would they result in inefficient networks, as we might predict based on the famous inefficiency of the waggle dance? Such studies could provide mutually valuable insights into both bee biology and bio-inspired computing.

Finally, we must acknowledge that our results reflect recruitment networks as they formed under our particular, experimentally controlled conditions. Recruitment network outcomes might plausibly be affected by any number of other factors that influence foraging and waggle dance behavior, including the concentration of the sucrose reward (Seeley et al., 2000), colony size (Eckert et al., 1994; Beekman et al., 2004; Donaldson-Matasci et al., 2013; Snyder et al., 2024), composition of the surrounding landscape (Seeley, 1995; Ohlinger et al., 2024), time of year (Carr-Markell et al., 2020; Steele et al., 1995), a colony’s nutritional status and needs (Seeley, 1995; Ohlinger et al., 2022a), and the presence of scent cues (Tautz, 2022) may well affect network outcomes. Additionally, this study gives results from fewer replicates than are typical in this field (two instead of three) due to the extreme labor demands of creating fully-marked colonies, and should therefore be considered an opening foray into the study of the waggle dance recruitment network. Future work to elucidate differences and similarities between recruitment networks in these different biological contexts, and with greater replication, will provide fruitful insights into the role of social networks in the lives of bees.