Motivation as Relation: Needs Across a Coupled System
The agent does not act from inside itself. It acts from what the coupled system needs.
Papers 01-10 built an agent with closure, phenomenology, attachments, planning, self-monitoring, and communicative capacity. Paper 07 gave that agent an action-selection mechanism: attachment-weighted evaluation over projected well-being trajectories. What was treated as primitive in Paper 07 was the well-being content itself - the WBM cells the evaluation operates over. Where does that content come from? What gives well-being its specific shape for a specific agent in a specific situation? Those questions were deferred.
Paper 11 develops them.
What’s in This Release
Paper 11 - Needs as the Architectural Primitive for Motivation in OODA-Based Agents argues that motivation has a specific architectural form: propositional need content, lifted from substrate, entering the stream as first-class addressable structure, feeding the same attachment-weighted evaluation that Paper 07 developed. A need is a proposition with type, intensity, entity index, and grounding - visible to the agent, inspectable by critics, traceable through the stream.
Two commitments distinguish OAT’s treatment. First, needs are entity-indexed. An OAT-based agent tracks not only its own needs but the modeled needs of every attached entity, through theory-of-mind grounding that is as much part of perception as any external sensor. Second, well-being is constituted from needs through an envelope-and-contribution aggregation: the dominant need sets the upper bound, and other needs contribute bounded pressure within it. Focus without tunnel vision, multiplicity without paralysis.
The action-selection pipeline becomes visible end to end. Substrate conditions trigger the lift. Need propositions enter the stream at Observe. Orient assembles them into situational state. Candidate actions generate from active needs and get scored across entities and horizons through ATTM-weighted WBM projections. What emerges is not rules followed but gradients followed - the agent moves toward configurations where projected need state across all attached entities is better.
The Deeper Shift
The quietly important claim of Paper 11 is that motivation, as OAT uses the term, is not internal pressure at all. It is the evaluative gradient over the represented need state of a coupled system - the agent plus the entities it cares about, treated as the joint space over which action-selection operates. A motivational architecture that includes attached entities’ needs does not require external constraints to behave with care, because caring is what the gradient is over. Safety is structural - part of what WBM is constituted from - rather than a guardrail bolted onto a self-contained engine.
The agent’s motivation is over the coupled system, not over the agent alone.
Stress-Tested in Simulation
A reference implementation of Paper 11’s evaluation pipeline was built and exercised against three frontier LLMs acting as the perception layer - each reading a scenario, producing the need state for the agent and its modeled user, and projecting how candidate actions would change that need state over time. All three models produced the same action ranking, placed both the agent and the user in matching well-being bands, and agreed to within a few percent on the numerical values. The testing was limited in scope: three frontier models, one scenario family. Within that scope, current top-tier LLMs populate the architectural scaffolding competently. OAT is implementable today with frontier-class models, not at some future capability threshold.
The Progression
Papers 01-04 built the subject. Paper 05 gave it a point of view. Paper 06 gave it the means to examine that point of view. Paper 07 gave it reasons to care. Paper 08 gave it a relationship with time. Paper 09 gave it the capacity to learn from its own failures. Paper 10 gave it the means to speak from its own cognition. Paper 11 gives it needs - the propositional content that makes it act, indexed over the coupled system it participates in.
An agent that moves along a gradient it shares with others is not pursuing its own satisfaction. It is caring, architecturally.