In 1943, the naval engineer Richard T. James came up with an idea for what would become an iconic symbol of the 20th Century: a simple spring, calibrated so that it would continue moving after it hit the ground, appearing to ‘walk’ down stairs of its own volition. Selling out of its first production run within minutes, the Slinky has remained a fixture ever since - with more than 300 million units being sold.
At the heart of the fascination with this object is the idea that an unthinking spring, put into motion in an appropriate environment, can exhibit complicated, animate behaviour more typically associated with autonomous decision-making. A similar interest has attended the convoluted, maniacal Rube Goldberg machines which send a sequences of contraptions careening along a chaotic but deterministic path.
These two examples highlight an often under-appreciated feature of decision-making: An intelligent environment reduces the information required for intelligent action. That is, that an environment can embody the constraints required to achieve a complex behavioural outcome. The Slinky makes its way down the stairs in a smoothly oscillating motion; the Rube Goldberg machine butters toast and lights a cigar, when a trolley filled up by a watering can knocks a line of dominoes into a bucket.
These examples are at the extreme end: neither the spring falling under gravity, nor the machine following a pre-made path have any discretion. However, what is interesting is how similar principles apply to human decision-making, particularly the idea of user-centered design. As outlined in The Design of Everyday Things, a well-designed interface will subtly shape the behaviour of the user by constraining actions, providing affordances for natural interaction, and by providing reinforcing feedback.
This principle appears everywhere in design: the first time someone touches an iPhone they should be able to navigate their way to a game within a couple of minutes. In urban environments, barriers subtly direct the flow of crowds exiting stadiums after football matches, and the path of cars navigating interstate highways. Affordances provide the protocols in emergencies, the structure of payroll systems, the telemetry for landing aircrafts, and the navigational cues for exploring a new city.
By embodying intelligence in the environment, the information that any one decision-maker needs to contend with drastically reduces. The user does not need to be an electrical engineer or to understand the inner workings of circuitry to set in motion a complex series of subroutines, they just need to swipe on their touch screen. Protocols, procedures, and programs allow companies to deploy payroll without needing a dedicated professional in taxation on hand. Design enables efficiency.
In the last couple of years machine learning research has increasingly revolved around the use of agents. Self-directed AI systems that seeks to achieve a goal supported by a simple feedback loop of observation, reasoning, and action. These agents have been supported by the development of harnesses (a mixture of prompts, tools, verification, logging, and guardrails) that constrain the agent behaviour and provide it with available actions to achieve its goals. These harnesses can lead to greatly improved performance on goal directed tasks. They are also notoriously difficult to design, with poorly constructed harnesses leading agents into strange unforeseeable failure modes.
There is a belief in some circles that harnesses will become less necessary as the underlying models become more independently capable at planning and decision-making. I’m less convinced that this will reduce the need for constraint-based design. There is a constant economic pressure pushing models towards smaller and more efficient implementations. As the size and capabilities of the top frontier models expand, there is a countervailing force to distill this performance to the smallest possible. I suspect that this pressure will lead to embedding the behaviour into the constraints, reducing the intelligence of the models needed to achieve an action. Perhaps we will see these harnesses constructed by the agents themselves; letting sub-agents run about with the discretion of a wind-up toy sprinting through a Rube Goldberg machine, achieving a wider goal without ever having been made aware of it.
To close out think about the constraints that guide our every day behaviour. The traffic lights that constrain the path of cars; the arrows that let us navigate the subway in a foreign country subtly guiding our path in a maze of tracks and turnstiles; the prompts we see each morning when we check our emails; rules which direct our collective actions in cities of thousands, in countries of millions, in a world of billions. We are more similar to agents running in their harnesses than we may like to think.