Division of labor is central to collaboration: Even young children can appropriately assign tasks based on the relative competence of their collaborators (Baer and Odic, 2022, Magid et al., 2018), or take their collaborator’s physical constraints into account when planning their own actions (Warneken et al., 2014). One of the chief advantages of division of labor is that it allows individual workers to get better at specialized tasks through practice. In The Wealth of Nations, Smith (1937) gives the example of a contest between a teenager and a smith: Through sheer practice, a teenager who solely makes nails, day in and day out, can vastly outperform a smith who only occasionally makes nails as one of the many products of her forge. Smith (1937) argued that these increases in efficiency from division of labor could directly account for extraordinary increases in the wealth of nations. Indeed, many of the benefits of collaboration and division of labor stem precisely from the fact that competence is not static.

Taking into account how people’s competence changes with practice is particularly important when making decisions about whom to recruit for a collaborative task, or how much to invest in their training. For example, a company might hire a less experienced candidate who – with appropriate training – could qualify for the job, as opposed to a highly-skilled candidate who requires a hefty salary. Similarly, graduate school advisors might admit a student who shows great potential, given opportunities to train and learn. Training also allows current team members to become better suited for their roles. Organizations investing in effective training programs that enhance employees’ capabilities tend to achieve both short- and long-term benefits for both the employees and the organizations (Nda & Fard, 2013). Effective training fills the gap between desired performance and actual employee performance; by improving employees’ performance and capabilities, companies increase their organizational productivity (Elnaga & Imran, 2013). Formal employee training programs are also unique in their ability to bring below-average firms up to the performance level of comparable businesses (Bartel, 1994). In short, people often make decisions in real-world collaborations by considering not only how capable people are already, but also how capable they will become over the course of a long-term collaboration. However, little is known about the lay intuitions behind how people make decisions about whom to recruit, whom to train, and how to train them.

Answers to these questions hinge on many factors, including what we know about the team structure, what training resources are available, how much time we have for training, etc., and also require that we take into account how our collaborators’ competence may change on the job. These decisions can be characterized as a sequential decision problem that requires long-term prospection. Recent work suggests that – much as people can use planning to solve sequential decision problems (Daw and Dayan, 2014, Huys et al., 2015) in non-social settings – they can also plan joint actions efficiently (Curioni, 2022, Török et al., 2019, Török et al., 2021) and use intuitive theories of other people’s minds to plan interventions on their mental states (Ho et al., 2022), including beliefs and desires (Baker et al., 2017, Baker et al., 2009, Jara-Ettinger et al., 2016, Premack and Woodruff, 1978), and also representations of competence and effort (Xiang et al., 2023b). Thus, one possibility is that people may make decisions about whom to recruit and train for a collaborative task by anticipating how others’ competence may change over time and planning accordingly. However, planning in general is computationally expensive, and these costs may be prohibitive in the context of collaboration, where optimal planning requires recursive theory of mind (Vélez & Gweon, 2021).

Thus, people may instead use simpler heuristics to decide whom to assign to what task without planning ahead (Dhami, 2003, Payne et al., 1993, Payne et al., 1996). For example, people may simply assign tasks based on who is more competent now, without considering how their competence will change as a result of training. Indeed, past work has found that adults tend to select teams whose combined expertise suffices for a particular job (Xiang et al., 2023b), and that children tend to use relative differences in ability to assign people to tasks (Baer and Odic, 2022, Magid et al., 2018). Another possibility is that – rather than treating training as a means to an end – people might instead train agents who need it the most. If this is the case, then they may selectively train the weakest agent, or the agent who stands to improve the most from a single bout of training, without planning ahead to consider what the group’s overall competence will be after training.

Across four experiments (N = 396), we show that people engage in planning when they make training and hiring decisions regarding simulated agents. The first two experiments are a collaborative box-lifting task framed as a game show. In Experiment 1, participants trained agents to prepare them to lift a heavy box together. In Experiment 2, participants first selected two agents out of a pool of four candidates, then trained the selected agents to lift a heavy box together. Experiments 3 and 4 generalized the task to a more education-oriented context, where participants recruited and trained students for a math Olympiad. We compared the human judgments to a planning model and four alternative models that base training decisions on heuristics: an exploitation model that selects the strongest agent, an equity model that selects the weakest agent, a learning model that trains agents who will improve the most after a single training step, and an equality model that invests equally in every agent’s training. We found that the planning model qualitatively and quantitatively provided the best match to the data.1 In the following section, we describe these models in detail.