Why AI Sourcing Agents Are Only as Smart as Their Sensors and Actuators

A brilliant algorithm paired with poor data or weak execution will underperform a simpler one paired with richer inputs and stronger delivery, true for robotics, multi-agent swarms, and sourcing and negotiation alike. Real intelligence isn't a standalone brain; it's perception, reasoning, and the capacity to carry that reasoning out, working as one.

Key takeaways

• Intelligence in multi-agent systems can't be separated from the sensors that perceive the world and the actuators executing it
• The same logic applies to sourcing: better data inputs ("sensors") and richer negotiation mechanisms ("actuators") directly improve outcomes
• Multi-agent coordination depends on shared perception between agents (stigmergy) and the ability to adapt gracefully when one agent's capability degrades
• Systems built with intelligence, sensing, and action co-designed from the start outperform those bolted together in silos

"You don't have to be an engineer to be a racing driver, but you do have to have Mechanical Sympathy." — Sir Jackie Stewart

When three-time Formula One World Champion Jackie Stewart coined the term "Mechanical Sympathy," he wasn't talking about treating machines with emotional warmth. He was describing a profound, tactile understanding of the physical machine he was piloting. Stewart knew that optimal performance didn't come from simply mashing the throttle and expecting the car to obey. It came from sensing and understanding the raw capabilities, the inherent friction, and the absolute limits of the engine, the transmission, and the tires given the prevailing conditions. The driver and the machine had to operate as a single, harmonious unit. To push a car to its limit without breaking it required a mind perfectly attuned to the physical reality of the vehicle.

Decades later, software engineer Martin Thompson applied this concept to computer science, arguing that writing highly performant code requires understanding the underlying hardware: CPU caches, memory bandwidth, and network latency.

Today, as we stand on the frontier of complex, autonomous Multi-Agent Systems (MAS): from swarms of delivery drones to networks of algorithmic trading agents: Jackie Stewart's philosophy has never been more critical. Industry-wide, over 92% of sourcing events are operated by agents either fully autonomously or with minimal human in the loop intervention. So B2B commerce is evolving rapidly towards multi agent systems. We are (re)discovering a hard truth: in the realm of autonomous systems, intelligence cannot exist in a vacuum. It is fundamentally, inextricably bound to the sensors that perceive the world and the actuators that manipulate it. This is important because in Procurement, sourcing optimization has more powerful sensors and actuators so the effective intelligence and performance is greatly enhanced relative to base workflow processes. It's akin to Jackie Stewart driving a F1 racing car versus a bicycle. Here is an exploration of why true intelligence in multi-agent systems is not a disembodied algorithm, but an emergent property of the mind, the senses, and the physical capacity for action.

The Myth of the Disembodied Brain

Why does AI intelligence fail without an understanding of its own sensors and actuators?

There is a pervasive fallacy in how we often discuss artificial intelligence. We tend to conceptualize AI as a "brain in a jar"—a pristine, abstract neural network that crunches data and spits out perfect decisions. There is a proliferation of startups professing to have multi-agent systems that read in tokens and output tokens without any appreciation of the semantics of the interconnected systems. There is an abundance of LinkedIn posts on AI and parameter counts, logic layers, and training algorithms, treating the physical body of the system as an afterthought.

This view falls apart when an agent is forced to operate in the physical world or within highly constrained digital environments. When a robotic agent needs to navigate a cluttered warehouse, or a software agent needs to execute high-frequency trades across congested networks, the "brain" is only as good as its connection to reality.

In a multi-agent system, an agent's intelligence is defined not just by its capacity to reason, but by its capacity to interact. If an algorithm calculates a flawless trajectory for a robotic arm, but the physical servos lack the torque to execute the movement in time, the system fails. The intelligence of the system wasn't "smart enough" because it lacked Mechanical Sympathy; it failed to account for the boundaries of its own physical actuation. Intelligence, therefore, must be embodied.

Sensors: The Texture of an Agent's Reality

What role do sensors play in how an AI agent understands its environment?

To an autonomous agent, the world does not exist as an objective, perfectly clear reality. The world only exists as an interpretation of sensor data.

Whether these sensors are physical (LiDAR, RGB cameras, piezoelectric tactile sensors) or digital (API endpoints, network packet sniffers, weather data streams, financial data etc), they are the agent's sole window into existence. And sensors are inherently messy. They suffer from:

• Latency: The delay between an event happening and the sensor registering it.
• Noise: Random fluctuations that obscure the true signal.
• Blind Spots: Physical or digital occlusions where the sensor simply cannot see.
• Bandwidth Limits: The maximum rate at which data can be ingested.

Mechanical Sympathy in the context of sensors means that the intelligence layer must implicitly understand these limitations. A smart agent doesn't just read a LiDAR point cloud; it understands the statistical probability of a false positive caused by a speck of dust or a raindrop.

In a multi-agent system, this becomes exponentially more complex. Agents must not only interpret their own noisy sensor data, but they must also communicate and synthesize this data with other agents. If Drone A detects an obstacle, but Drone B's sensors read clear airspace, the collective intelligence of the swarm must resolve this conflict. The algorithms driving the system must be hard-wired to respect the refresh rates and error margins of their specific sensory hardware. You cannot hot-swap a cheap camera for a high-end LiDAR system and expect the same control algorithm to function optimally; the "brain" must be uniquely tuned to the "eyes."

Now consider this in the context of sourcing and supply chains; the sensors or inputs are:

• Unstructured, ambiguous requests from a multiplicity of sources across many spend categories.
• Urgent versus non-urgent cases
• Unmanaged or maverick spend detected after it occurs
• Capacity limits on personnel in Procurement
• Suppliers that struggle to articulate or express complexity in their cost structures
• Misunderstandings on the part of suppliers and/or stakeholders.

Intelligence in the context of Procurement relies upon extrapolating or interpolating based upon available data and reasoning about gaps or inaccuracies. Improved sensors help matters and an intelligent system should be good at applying better practices to close information gaps and achieve better outcomes.

Actuators: The Friction of Action

Why does an AI agent's ability to act matter as much as its ability to reason?

If sensors define what an agent can perceive, actuators define what an agent can do. Actuators are the mechanisms of intervention—the motors, the servos, the hydraulic pumps, or, in software, the functions that write to a database or execute a trade.

Just like sensors, actuators are bound by the laws of physics and system constraints:

• Torque and Force: The maximum physical effort an actuator can exert.
• Slew Rate: How fast a system can transition from one state to another.
• Backlash: The mechanical "play" or looseness in gears and joints.
• Thermal Limits: The point at which a motor overheats or degrades if pushed too hard.

A disembodied AI might calculate that to catch a falling object, a robotic arm needs to accelerate instantly to 50 meters per second. A system designed with Mechanical Sympathy knows that attempting this will strip the gears or burn out the motor. Therefore, the intelligence of the system must be co-designed with the actuators. The control logic (like PID controllers or model predictive control) relies heavily on understanding the mechanical friction, the mass of the moving parts, and the energy required to stop them. In a multi-agent environment, where robots might be collaborating to lift a heavy object, the precision of their actuators must be perfectly synchronized. Intelligence isn't just knowing where to move the object; it's understanding the exact micro-adjustments in torque required by each individual motor to prevent the object from dropping or tearing apart.

Likewise, when sourcing goods and services there are sensors and actuators to consider. If the most intelligent negotiation strategy is to invite expressive bids from suppliers in a synchronous multi-round negotiations, with feedback loops, and scenario optimization, but the negotiation is attempted over email and lacks the 'sensors' to gather rich data and the 'actuators' to offer feedback and reason about complex trade-offs then you get a poor outcome. Intelligent agents require powerful sensory and actuation capabilities to have effective intelligence. Tokens that express what you should do are of little real value if they cannot deliver action and results.

Symbiosis in a Multi-Agent System World

How do agents coordinate perception and action when they share an environment?

When we scale up from a single agent to a multi-agent system (MAS), the inseparability of intelligence, sensors, and actuators becomes the defining characteristic of success or failure. As Professor Mike Wooldridge remarked at ProcureTEX 2025 in London last October: in a MAS, agents share an environment. The actions of one agent immediately become the sensory input for another. Consider a swarm of autonomous vehicles navigating a complex intersection without traffic lights.

• Coupled Perception: Car A's sensors must detect Car B. But Car A's intelligence must also predict Car B's limitations based on its current trajectory.
• Stigmergy: Agents often communicate indirectly by modifying the environment—a concept known as stigmergy. In warehousing, a robot moving a pallet (actuation) changes the layout of the warehouse, which must instantly be perceived (sensing) by all other robots to update their pathfinding (intelligence). In sourcing, the world for bidding agents changes when demand is updated, requirements change, or feedback on their competitiveness is communicated.
• Cascading Constraints: If one agent's actuator degrades (e.g., a drone loses a rotor and has to fly slower), the intelligence of the entire swarm must adapt to this new physical reality. The swarm's routing algorithms must route around the slower agent, adjusting their own actuation speeds to maintain formation. Similarly in Procurement, as a demand forecasting agent signals an increase in expected demand, it alters the input and thresholds for what may be considered sufficient competitive tension in a negotiation.

In these systems, you cannot draw a clean line between where the real world ends and the intelligent software system begins. The limitations of the actuators dictate the required sampling rate of the sensors, which in turn dictates the processing power and algorithmic architecture of the intelligence layer. In Procurement settings, one should assume sourcing agents are interacting with bidding agents so this coupled perception is important for overall system performance.

Co-Design: Building for Sympathy

How do you design intelligence, sensing, and action to work in harmony?

So, how do we apply Jackie Stewart's wisdom to the future of AI? We should accept the reality of our world: a siloed approach to engineering intelligent systems doesn't work well. Just as we cannot have a hardware team build a robot, hand it off to a software team to write an operating system, and then hand it to an AI team to "make it smart." So too we cannot have a software system handed to an "intelligence" team to build controller agents. Intelligent systems requires co-design. Keelvar Systems comes from a Gaelic word meaning intelligent systems: we built a sourcing optimization workflow in harmony with an agentic layer that embeds intelligence in sympathy with the sensors and actuators. These sensors were deliberately designed to be ultra powerful, more powerful than most human users could apply in regular sourcing. But agents thrive on more data and richer sensory input; they can quickly learn how to apply the powerful features in optimization to even the smallest sourcing events.

When we build multi-agent systems, we must design the software architecture to gracefully handle the inevitable degradation of physical parts or uncertainty of input data. We must build AI that doesn't just issue commands, but listens to the feedback of the system, understanding when errors or omissions are prevalent. Intelligence is not a magic ghost you can plug into any machine. It is the emergent harmony of perception, computation, and action. To achieve optimal performance in the complex, unpredictable world of multi-agent systems, we must recognize that the mind, the senses, and the body are one. We must design with Mechanical Sympathy.

Credit: This article was inspired by a discussion with Keelvar's CPO, David Devlin, who pointed to Jackie Stewart's point about mechanical sympathy and its relevance to intelligent systems.